AI Trending Rankings

Continue to No.04–10

1. 04
   Claude identifies the source of NIST's elliptic-curve seed

   Anthropic said its internal research model Claude identified, in less than half an hour, the source of the seeds for the five standard elliptic curves NIST P-192, P-224, P-256, P-384, and P-521: These seeds were not hash results, but outputs from the C library’s lrand48 random-number generator initialized with Unix timestamps. For example, P-256 requires 1,842 iterations. The team has publicly released reproduction code. Anthropic stressed that this was not a break of ECDSA, and that no new attack method or curve weakness was discovered, so users do not need to take any action as a result. The $36,864 prize associated with the discovery will be donated to the Internet Security Research Group. According to The Decoder, these curves were standardized in 2000, and their seeds were provided by the NSA with their origins unclear, fueling suspicions of a backdoor for years. X user Yuchen Jin called it a cryptographic puzzle left unresolved for 25 years.

   AhQuantity+882Unchanged
2. 05
   NVIDIA RTX Spark laptops go on sale, with conflicting reports on starting prices

   NVIDIA’s RTX Spark laptops are now on sale, launched by ASUS, Dell, HP, Lenovo, Microsoft, and MSI. They can run local coding agents on a single machine to handle coding, testing, and deployment. In its launch-day press release, NVIDIA said prices start at $2,599 and included a limited-time offer for new Perplexity users; an October 18 blog post, however, said prices start at $2,399. NVIDIA said pricing and configurations vary by region, and that compact desktop systems will arrive in November. The product is based on the GB10 SoC, with up to 128GB of unified memory and 1 petaflop of FP4 compute. Three editors at The Verge spent a week testing five Windows on Arm laptops powered by the RTX Spark N1X chip, priced from $2,599.99 to $6,999.99. X user Kazike said the ASUS ProArt 16-inch laptop with 128G+1T costs 37,999 yuan and sold out online almost instantly; in his tests, Qwen3.8-27B ran at about 30 tokens per second, while a local 2B translation model supported six concurrent sessions. OpenClaw says it now supports out-of-the-box setup on the new laptops, allowing users to choose local…

   +1070↓ 21%
3. 06
   Google releases Gemini 4 Argon, available exclusively to paid users

   Google has officially released its flagship model, Gemini 4 Argon, making it available worldwide starting today. It is designed for complex coding, agentic, mathematical, and scientific tasks. The model is available only to paying users: developers and enterprises must use paid tiers of the Gemini API and Gemini platform; in Google Antigravity, it is limited to Ultra subscribers; and in the Gemini app and Search AI Mode, it is limited to Ultra and Pro subscribers. The Decoder says free users and Plus subscribers can only continue using the weaker Flash series models. As for pricing, Google announced launch-period rates of $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached inputs. MarkTechPost says that after November 30, 2026, prices will rise to $4 for input and $20 for output. The model supports a unified input-output limit of 1M tokens. Google also said that scientists and researchers can apply for six months of free access to Antigravity’s highest usage tier, while cybersecurity defenders can apply through the Fairwind Program; both are subject to eligibility, capacity, and screening requirements. Googl…

   DG+1370Unchanged
4. 07
   Tencent Hunyuan released and open-sourced Hy Memory 2.0

   Tencent Hunyuan released and open-sourced the agent memory framework Hy Memory 2.0 on October 16, offering three layers of capabilities: the dedicated open-source model Hy-Memory-30B-A3B (30B total parameters, approximately 3B active parameters in an MoE), the programmable memory framework hy-memory SDK, and the desktop tool MemHub. The three can be used independently or combined into a complete Memory solution, positioned as a personal memory assistant. Tencent Hunyuan said the framework handles three types of memory: chat, work, and code. It ranked first among open-source offerings on four benchmarks, while Hy-Memory-30B-A3B surpassed Opus 5 on the major frameworks. MemHub provides a CLI and a Mac version and can synchronize across more than 10 AI tools. The SDK and model are available on GitHub and Hugging Face. The above evaluation results and performance claims are from Tencent Hunyuan.

   PublicD+868↑ 15%
5. 08
   OpenAI brings computer use capabilities to Chrome

   OpenAI upgrades ChatGPT’s Chrome extension, bringing computer use capabilities to the browser. In Chat mode, ChatGPT can read users’ open tabs and summarize information; in Work mode, it can control tabs to send messages, fill out forms, download files, and organize tab groups. The Verge says you can invoke it by typing @Chrome in a web or desktop app. The extension is now available for download from the Chrome Web Store, and ChatGPT says it is available on all plans. IT Home reports that no paid subscription is required, and that it also works with Chromium-based browsers such as Microsoft Edge, Brave, and Opera. IT Home also says Work mode requires authorization to control tabs and, by default, asks for confirmation before accessing each new website. The Decoder says GPT-6 Astra is currently the most powerful Computer Use model.

   +1168Unchanged
6. 09
   Anthropic’s third-quarter revenue reaches $16.8 billion as growth slows

   Anthropic disclosed at a closed-door investor meeting in San Francisco that its third-quarter revenue was $16.8 billion, up approximately 45% quarter over quarter to nearly 50%, a significant slowdown from the 147% growth rate in the second quarter. The company also said that every $1 spent on computing power generated $1.70 in revenue, with an adjusted operating margin of 11% for the year and positive cash flow. Some investors were disappointed by the revenue figures and expected the potential IPO valuation to fall below $1.5 trillion, down from the previously discussed $2 trillion. Reuters, citing people familiar with the matter, said the Investor Day event discussed a valuation of approximately $1.5 trillion, while bankers' rumored valuation of around $2 trillion and fundraising of more than $10 billion had not yet been finalized. Bloomberg noted that the market interpreted the slowdown in growth as bad news for AI chips and servers. Sources gave differing accounts of second-quarter revenue: some reported approximately $11.5 billion, while others said it had risen from approximately $4.8 billion to $11.5 billion. The figures regarding $65 billion in ARR in July and revenue estimates for August and September came from independent calculations by X user Oran Ge, not from company disclosures.

   TitaniumPhoenixEG+959↓ 22%
7. 10
   Reflection’s open-source 501B model Beam launches on Hugging Face

   Reflection AI has uploaded the weights of its first open-weight model, Beam, to Hugging Face (reflectionai/Beam-1.0-501B-A23B) under the Apache 2.0 license. Previously, the company had only said that the weights, technical report, and model card would be released this month, and that early access would require joining a waitlist on the Reflection platform. The Decoder reported at the time that the company was still conducting final safety tests and offering early versions only to some users. Beam is a text-only sparse MoE with 501B total parameters and 23B activated per token. It was pretrained on 23.8 trillion tokens, supports a 1M-token context window, and targets coding, reasoning, and agent tasks. It offers BF16, FP8, and NVFP4 weights, as well as a speculative decoding drafter. SGLang says it supports Beam natively and powers Reflection’s production API; vLLM says it has Day 0 support and is optimized for NVIDIA Blackwell. Ollama says Beam is available for a limited time at no cost on Ollama Cloud, with zero data retention. Claims about its actual performance vary: Reflection says it matches GLM 5.2 on reasoning tasks with three to four times less compute, while its coding and agent benchmarks are close to Qwe…

   +957↓ 19%