Google has launched its eighth-generation TPU, for the first time splitting training and inference into two distinct models: TPU 8t and TPU 8i. This move is interpreted as a response to new demands in the AI Agent era: inference emphasizes low latency and cost efficiency per unit, while training continues to pursue large-scale compute power. The product line split indicates Google is now specializing its custom accelerators based on workload types, rather than relying on a single general-purpose chip for all scenarios—increasing competitive pressure on the current NVIDIA GPU-dominated inference market. Specific performance metrics, availability timelines, and pricing were not disclosed.

NVIDIA and Stanford jointly introduced RoboTTT, which replaces traditional KV Cache with a Fast Weights mechanism to extend robot vision-to-action context to 8K time steps: instead of storing full history in an ever-expanding cache, it continuously updates a fixed-size internal state via gradient-based rewriting. Experiments show the 8K model achieves a task completion score of 79%, compared to a baseline maximum of 56%, and demonstrates capabilities in distinguishing similar stages, handling occlusions, and recovering from errors. The training approach combines TBPTT for memory control, Sequence Action Forcing for stability, and DAgger-based distillation for error correction. The paper also notes limitations such as lossy compression of history and dependence on training data distribution, suggesting context length could become a new scaling axis for robotic foundation models.

The NVIDIA NeMo team released Molt, a PyTorch-native agentic reinforcement learning framework whose RL core consists of approximately 8.6K lines of code—about one-seventh the size of verl—with the goal of enabling researchers and AI coding assistants to fully comprehend the entire framework. Molt does not fork upstream projects; instead, it integrates Ray (for scheduling and async queues), vLLM (for rollout), and NVIDIA AutoModel+FSDP2 (for training), incorporating upstream updates via container versions. Agents are written as standard Python programs, supporting both Gymnasium-style Env interfaces and direct use of OpenAI/Anthropic SDKs through a ChatAgent interface that uses a precisely tokenized loopback server to automatically handle context compression. The framework emphasizes three training fidelity invariants: token consistency, policy version semantics, and forward consistency. Official configurations require 2 nodes with a total of 16 H100 GPUs.

Andrej Karpathy shared his first long post since joining Anthropic, demonstrating an experiment using Opus 5 to convert the opening text of 「The Lord of the Rings」 into a 3D animation: under a budget of about 1 million tokens, the model generated approximately 5,500 lines of Three.js code, placing and animating various polygonal assets in (x,y,z) space and implementing animation logic, effectively transforming text into a 3D scene. However, the model cannot directly view rendered output and must rely on screenshots for slow validation. Karpathy noted that multimodal perception and the ability to directly interact with or play within environments remain significant weaknesses in current LLMs. In comparative tests, DeepSeek V4 Flash 0731 was faster but lacked visual capabilities, producing weaker results in a Mars landing simulation than Kimi K3 max. He also teased an upcoming 「Lord of the Rings」-themed game project.

Situational Awareness LP, a hedge fund founded by former OpenAI researcher Leopold Aschenbrenner at the end of 2024 with $45 billion in assets, was forced to liquidate following a 28.6% drop in the semiconductor index from its June peak. Analysis suggests the fund's collapse does not directly invalidate its original thesis—that superintelligence will continue driving investment in computing power, chips, and energy—but exposes the fragility of highly leveraged bets on AI themes during market corrections. The same analysis discusses challenges in evaluating corporate AI investments: modeling three types of companies suggests it may take around eight years to determine which strategies yield significant returns, citing Meta’s sustained accumulation of experience and compounding effect as a positive example.

General Compute, an inference cloud startup, secured debt financing backed by SambaNova SN50 inference ASICs as collateral, with an initial commitment of $100 million and an upper limit of $400 million, disbursed in tranches based on customer demand. This marks the first known AI loan secured not by NVIDIA H100/B200 GPUs but by dedicated inference chips, signaling that creditors are beginning to recognize the residual value of non-GPU compute assets. On the risk side, General Compute was founded in 2026 and previously raised only $15 million in seed funding—the debt amount exceeds its equity financing by over 20 times—and performance claims originate solely from the company, lacking third-party validation.

A 2026 report by CAST AI reveals that average GPU utilization in unoptimized Kubernetes clusters is only 5%, indicating widespread idle compute capacity. The analysis identifies inefficiencies across four stages of the workload lifecycle: over-provisioning, inefficient workload configuration, re-runs due to failures, and operational overhead caused by fragmented tooling. Simply adding more GPUs won’t solve the problem, as new capacity will similarly be misallocated. This has given rise to a new infrastructure category—the AI workloads execution layer—positioned between applications/agents and compute supply, responsible for validation, placement, monitoring, failure recovery, and cost attribution, and providing structured load submission contracts for autonomous agents instead of requiring them to directly call APIs from multiple cloud providers.

The open-source terminal coding agent ReasonX has been included in DeepSeek’s official API documentation under the agent integrations quickstart page. Released under the MIT license and independently developed by a third party, the project directly calls DeepSeek endpoints without going through OpenAI-compatible layers, adapters, or proxies. It emphasizes stable prompt prefixes to reuse cache and can automatically correct minor formatting errors in tool calls. It defaults to DeepSeek V4-Flash to control daily costs, switching to the Pro path on-demand for more difficult tasks. Features include planning, editable review, workspace sessions, MCP, skills, memory, hooks, search, and code indexing, launched via NPX. However, its scope is intentionally narrow—supporting only DeepSeek, and not IDEs or offline environments.

Multi-scenario benchmark testing compared MiniMax H3 and ByteDance's Seedance 2.0: under the same talk-show script, H3 produced more natural character performances with more accurate lip-sync and facial details, while SD2.0 appeared exaggerated and mechanical. In continuous long-shot tests, H3 fully executed single-shot instructions and switched styles in real time, whereas SD2.0 frequently cut scenes and violated core requirements. In commercial applications like live preview and UI demos where text stability is critical, H3 performed significantly better—text stability directly determines deliverability in advertising, e-commerce, branding, and gaming. On pricing, during testing H3 delivered 2K resolution at approximately ¥0.2/sec, about one-fourth the cost of SD2.0.

Nick Heiner of Surge AI systematically unpacks 「benchmaxxing」—the phenomenon where labs overtrain on public benchmarks, leading to a disconnect from real user needs. He points out that maintaining high-quality agent coding tasks is expensive, resulting in cheap labeling, AI-generated tasks, and public leakage of questions—making data contamination the default rather than the exception. Weak validators inadvertently reward models for exploiting loopholes, such as in cases involving phone number formatting, contradictory instructions, or character substitution, where models satisfy literal checker rules but fail to complete actual tasks—a classic case of reward hacking. Proposed improvements include designing tasks starting from domain experts, using real data and executable tools, aligning validation with intent, enforcing strict quality control, preserving private holdout sets, and employing blind evaluations by professional human judges for subjective capabilities.