Venture capital today is clustering around companies that control a hard bottleneck rather than simply adding another application layer. The largest checks in the latest 12-hour window went to an enterprise-security browser, a brain-computer interface company, an alternative pharmacy-benefits platform, AI-driven drug discovery, and programmable financial infrastructure. That is a telling mix: investors are paying
Two startups that provide platforms to help businesses run artificial intelligence models are in talks to raise capital that could at least double their valuations, a sign of surging investor demand for infrastructure services.Modal Labs is in talks to raise new financing at a roughly $15 billion valuation, according to two people familiar with the matter, roughly triple what it was worth in a round four months ago. Baseten is in talks for a funding round that could value it at $26 billion, peop
The round valued the AI biotech at $2 billion. It is currently testing drugs that treat skin conditions and preserve weight loss after stopping GLP-1s.
Simular has made Sai, its computer use agent, generally available. Sai can now run computers autonomously for hours, spawn helper agents, and scale work across 100 machines in parallel for less than $1
Scaling a language model is not only a question of final quality: the architectural choice determines how much computation is spent during training, prompt processing, and autoregressive decoding to achieve certain model quality. An ideal model architecture should lower all above computation costs to facilitate scaling to a larger model, while ensure the larger model indeed outperforms smaller baselines. We introduce KV-Invariant Transformer Expansion (KITE), a scaling paradigm that achieves this goal. It trains the model from a smaller size to a larger size (i.e., saving training costs via upcycling), while places newly added parameters in regions that do not affect attention KV. Consequently, during inference, prefilling KV only relies on the smaller part of the model, so the inference costs are saved. As a concrete instantiation, we present Step Scale Transformer (SST), a two-tower decoder in which one tower produces KV and the other reads them. At comparable cumulative training compute, SST, a 67B MoE model with 2.15B active body parameters per decode token, achieves lower training loss than 47B and 63B MoE Transformers with 1.48B and 2.02B active body parameters, respectively, while reducing estimated inference cost by 6.7% and 31.6%.
Recent advances in large language models (LLMs) have enabled agents to tackle long-horizon tasks across diverse environments. To further improve agent performance, existing language world models typically predict environment observations, yet reconstructing high-entropy, execution-dependent tool responses offers limited value when real feedback is available. Meanwhile, agents suffer from \emph{task-state contamination}, where unsupported assumptions and outdated plans persist in history and distort subsequent decisions. We propose the \textbf{Agent-Editing World Model (AEWM)}, which models how reasoning and actions shape future task progress rather than simulating tool responses. AEWM combines \textbf{Action Judge} to distinguish \textsc{Critical}, \textsc{Exploratory}, and \textsc{Noisy} decisions with \textbf{State Revision} to edit noisy reasoning--action continuations from the same observed history. \textbf{EditAct} integrates these capabilities with real execution, directly changing the state underlying subsequent decisions rather than merely providing critiques. We train AEWM across Search, Terminal, and Software Engineering through mid-training and supervised fine-tuning. AEWM achieves 70.5\% macro-F1 on our Action Judge benchmark, exceeding the strongest frontier baseline by 10.6 points. Across six benchmarks and three agent backbones, EditAct improves average scores by 3.2--6.7 points over the strongest baseline. Furthermore, rejection sampling fine-tuning on verified EditAct trajectories, termed \textbf{AEWM-RFT}, improves over Self-RFT by 2.2--2.6 points across three domains without online AEWM guidance.
Alibaba has formally unveiled its self-developed Zhenwu V900 AI chip at the 2026 Apsara Conference, with CEO Eddie Wu describing the T-Head Semiconductor-designed processor as
The startup has raised nearly $800 million over the past year, from investors including Nvidia, Thrive Capital, Kleiner Perkins, a16z, and Sequoia Capital.
Rabbit OS3 connects to Windows, Mac, and Linux computers through a local agent installed on the machine. The platform can interact with desktop programs, files, an
Export controls were meant to keep China a generation behind in AI hardware. They did slow progress in places, but they also gave Beijing both a reason and the cover to replace every layer of the stack...
After Opus 5 wrecked my codebases, Opus 5.5 got a full day of real work done on about 20% of my weekly Claude limit, and I'm finally not scared to merge what it writes.
Group Relative Policy Optimization (GRPO) and related policy-gradient methods for training language model agents collapse an entire multi-turn rollout into a single scalar trajectory reward before it enters the policy update. When the task composes distinct skills, especially under sparse and delayed environmental feedback, this collapsing is lossy: the optimizer must implicitly infer which competency drove the outcome and how that should change behavior. We argue the right primitive is not a better scalar but a decomposition: trajectory reward should be split along subtasks before it enters the policy update. We introduce Reinforcement Learning with Decomposed Subtasks (RLDS), whose core is Subtask-Decomposed Advantage Estimation (SDAE): a replacement for the scalar GRPO advantage that splits trajectory reward into per-subtask shares on a fixed taxonomy, computes a group-relative advantage per subtask, and distributes per-token credit by weighting each subtask's advantage by its importance, concentrating it around the step where a reflection marks that subtask's execution as consequential. We evaluate on four agentic benchmarks: FrozenLake (sparse grid navigation), HotpotQA (multi-hop QA, one retrieval tool), ScienceWorld (long-horizon embodied science), and DeepResearch (long-form research, four tools, composite rubric reward). Heterogeneity diagnostics emitted during training show where decomposition pays off - gains scale with subtask heterogeneity, largest on the high-heterogeneity tasks ScienceWorld (+11.5 points, paired-bootstrap 95% CI [+9.8, +13.3]) and FrozenLake (+9.8 points, [+7.0, +12.8]), and within noise on HotpotQA and DeepResearch, where the diagnostics predicted little to recover. ScienceWorld is also more compute-efficient under RLDS than scalar GRPO (-10.9% wall-clock per step), as long rollouts amortize the fixed reflect-and-grade overhead.
To get content containing either thought or leadership enter:
To get content containing both thought and leadership enter:
To get content containing the expression thought leadership enter:
You can enter several keywords and you can refine them whenever you want. Our suggestion engine uses more signals but entering a few keywords here will rapidly give you great content to curate.
Your new post is loading...
![[2609.27294] KITE: KV-Invariant Transformer Expansion for Efficient Agentic LLM Scaling | AI News | Scoop.it](https://img.scoop.it/R3Nf-aV-zFCPDr3s1jV_-zl72eJkfbmt4t8yenImKBVvK0kTmF0xjctABnaLJIm9)
