NVIDIA researchers proposed SEED, a method that uses shallower layers and lightweight emitters to write memories needed for later generation while retaining full-depth computation during decoding. On Qwen3.8-Flash-Next, parameter-dependent prefill computation fell about 25%, from 12.1 to 9.1 GFLOP per non-final input token. SEED also stores recurrent state in a compressed representation, reducing Qwen3.8-Flash-Next's per-request state from 113.2 MB to 29.2 MB immediately after compression and averaging 43.4 MB over the compression cycle. The report describes tests on five hybrid models and an arXiv preprint. In GB300 GPU and SGLang tests replaying coding-agent sessions, NVIDIA reported 16%–42% higher peak throughput per GPU; gains varied by deployment and workload.
TMTPost reports that Google co-founder Sergey Brin has resumed working with researchers on large-model architecture, while ByteDance founder Zhang Yiming has attended Seed core-team reviews regularly since the second half of 2024. The commentary frames both founders as re-entering technical leadership and cites a shared ambition to catch OpenAI and Anthropic by 2030, but it provides no primary confirmation of that target or several personnel details.
Nubia announced that the NaviX Ultra Starry Sky Collector's Edition opened preorders on November 10 at 10:00, with sales scheduled to begin on November 16 at 10:00. The special edition costs RMB 6,999 and is available only in a 16GB+512GB configuration, using a micro-nano relief process inspired by Van Gogh's “The Starry Night.” Nubia describes the phone as an AI agent device capable of executing cross-app tasks from a single command, with Doubao Phone Assistant leading its AI assistant experience.
A commentary recommends giving coding agents explicit goals, boundaries, acceptance criteria, repeatable checks, and evidence requirements while retaining human judgment over whether changes are deliverable. It says these practices apply to project-operating agents such as Claude Code and Codex, although exact commands may vary by tool and version. In a reported 15-run Seed table experiment, 5 of 10 runs using a table-configuration skill wrote “需要补货” into the existing listing-status column, illustrating that a completed workflow can still produce incorrect edits. Because each configuration was tested only five times, the result does not establish a general error rate.
IT之家, cnBeta and TMTPost report that ByteDance has introduced restrictions on using personally funded Claude and Codex accounts to access L4 code repositories and Feishu documents. A source cited by 36氪 says public work information and non-work personal use remain unaffected. ByteDance technical vice president Hong Dingkun attributed the change to the serious consequences of leaking L3-and-above content, while saying other tools should not be blocked merely to promote ByteDance products.
A Tech商业 report republished by 钛媒体, citing leaked internal emails and multiple reports, says ByteDance's Seed foundation-model team has completed a reorganisation under Wu Yonghui into five first-level departments: pretraining data, reinforcement learning, B-end/agentic post-training, C-end chat post-training, and safety and alignment. The new Seed Model-Alignment & Preparedness unit is led by Yan Lin and reports to Wu, with the former Responsible AI team and some Chat team members moving into it. The report also says evals lead Huang Wenhao has left and may start a company, and calls the lab China's first frontier AI lab with a standalone safety and alignment department; ByteDance has not publicly confirmed these details.
Nubia said its NaviX Ultra "beta lab" program will open soon, letting early adopters test the newest capabilities first, with the first evolution livestream scheduled for Oct 14 at 19:00. The device, called the "second-generation Doubao phone," is marketed by Nubia as the "world's first AI agent phone," built end-to-end by ZTE's Nubia, and pitched on letting users execute complex cross-app tasks with a single spoken instruction, with the Doubao phone assistant leading the AI experience; it starts at 5,999 yuan. Hardware includes a Snapdragon 8 Elite Gen 5 platform, a 6.78-inch 1–144Hz LTPO 2.0 display, a 200MP main camera, a 7,100mAh battery, a dedicated AI key with integrated fingerprint recognition, and Seed's latest full-duplex speech model.
An analysis of ByteDance's careers site counted 11,187 open China social-recruitment roles on Oct 5, of which 4,791 (42.8%) relate to AI — with AI's share of newly posted roles rising from 33% to 55%. By unit, the Seed model team was 96.1% AI roles, Doubao and other AI-native apps 77.2% and CapCut 74.3%, while e-commerce (26.4%) and local services (16.9%) trailed. The author argues ByteDance is hiring across six layers — chips, compute, foundation models, cloud, applications and devices — a "saturated, full-stack" push that comes at the cost of focus.
ByteDance's Seed team says in a new research paper that large language models' performance swings on very long inputs stem mainly from "phase sensitivity" introduced by chunked KV cache compression. To cut memory use in long-context inference, the method compresses consecutive token windows into fewer entries at a fixed stride, creating a new positional coordinate: a token's "phase" relative to the compression window boundary. Experiments show the same information becomes much harder or easier to retrieve at different phases, with long-context retrieval accuracy gaps of up to 40 percentage points in some large open-source models.
What problem it solves
Real-world translation involves more than converting text between languages: tasks can require terminology, style, personalization, delimiter preservation, and structured-data constraints. Hy-MT2 is designed to support translation among 33 languages and follow translation instructions in multiple languages for complex real-world scenarios.
Key features
Instruction-driven tas…
What problem it solves
Text-to-speech systems must generate expressive speech while supporting multiple languages and varied voice requirements. VoxCPM addresses multilingual synthesis, creation of new voices from descriptions, controllable cloning from reference audio, and continuation from an audio prompt and transcript.
Key features
Multilingual speech synthesis — Input text can be synthesized …
What problem it solves
Agents can retrieve documents and query databases, but unreliable results undermine agentic analytics. Teams need a context layer that records where data lives, what its definitions mean, whether it is usable, how it should be queried, its provenance, and related concepts.
Key features
Enterprise ontology generation — Generate semantic ontologies from database metadata, infe…
AI has already entered code writing, algorithm optimization, and model development. But after a task is completed, what remains may be reusable capability—or merely an answer and scattered logs. The difference becomes apparent only when a similar problem comes up again: can the system reuse what it learned, or must it pay the cost of exploration all over again?
Recursive Self-Improvement (RSI) co…