PHP officially released versions 8.5.12, 8.4.27, 8.3.36, and 8.2.35 on October 21, patching CVE-2026-109618 and CVE-2026-109619. OpenAI states its models participated with Patch the Planet researchers in discovering the vulnerability, and coordinated with maintainers and security vendors to prepare fixes in advance; this statement is part of OpenAI’s description of its own project. Bao Yu summarized that the vulnerability may allow unauthenticated requests to execute code, and reminded hosted users to confirm upgrades with their service providers. Cloudflare rules only provide temporary mitigation, and upgrading is the only complete fix. The earlier alert reposted by MaxForAI noted that the full scope of impact and deployment conditions were still awaiting official details at the time, so no overbroad claims should be made based on it.<\/p>\n
Source:<\/p>\n
DeepSeek officially announced it has open-sourced its internal inference framework DeepInfer, which it uses for large-scale deployment of V4.1-Flash, with goals of high throughput, low latency, and reduced computing power costs. The official project description reposted by Bao Yu lists capabilities including distributed inference, MoE parallelism, speculative decoding, and asynchronous scheduling, along with support for NVIDIA SM100 and Huawei Ascend 950, and an Apache 2.0 license. MaxForAI adds that the currently released version is for offline inference, and online KV Cache management and load balancing components have not all been made public yet. The value of this progress lies in the opening of production-grade inference engineering code, but it does not mean a complete online service stack has been released.<\/p>\n
Source:<\/p>\n
Bao Yu quoted remarks from Zongwei Zhou, Meta’s PyTorch-TPU lead, published on Google Developer Blog, stating that SGLang-TorchTPU is already running in Meta’s production environment. This is the first-hand confirmation available to date. This news confirms that TPUs are already used for real-world production inference, but public materials do not disclose deployment scale, specific products, or whether the chips are leased or purchased. Therefore, it cannot be inferred that Meta has fully replaced GPUs.<\/p>\n
Source:<\/p>\n
MaxForAI compiled this based on Reuters’ reporting on Anthropic’s draft prospectus: Anthropic posted $163 billion in revenue in the first half of this year, up from $11 billion in the same period last year. Of the $807 billion net loss, around $720 billion comes from the fair value increase of convertible financing instruments, which should not be directly counted as operating cash burn. MaxForAI also notes that computing and infrastructure spending reached nearly $90 billion in H1, Q3 revenue and adjusted profit data come from the company’s investor meeting. The over $6000 billion in planned future investment spans multiple years, and around $1500 billion in deals are still under negotiation. Plans for an IPO and valuation have not been finalized, and should be distinguished from realized operating data.<\/p>\n
Source:<\/p>\n
GACC-Bench, introduced by MaxForAI, is designed for agent harnesses and memory systems: when a model compresses context, it only receives the target and token budget, then uses the same compressed result to answer multiple unannounced questions under that target to test whether key information is retained. The benchmark includes 7 sources, 281 contexts, 500 targets, and 2571 questions, with samples expert-audited. Its report finds that when retaining 10% of the original tokens, the best model can answer all related questions correctly for only 14.8% of targets, compared to 96.8% for full context. This is a strict all-question accuracy metric, and the results suggest that evaluating compression should not only focus on summary fluency, as the target itself also affects what content is retained.<\/p>\n
Source:<\/p>\n