
Know AI. See What’s Next.
Slide 1 of 5: 姜宇辉的AI分身:知识助手与“另一个自己”之间
Latest news

HolyMolly links AI-generated 3D assets with game creation and interactive communities (opens in a new tab)
The article follows Tripo's move from generating assets to building games, covering text-based creation, Remix and examples. It also examines interaction continuity, gameplay and runtime quality as the platform explores an interactive-content community.

Comparing valuations and annualized revenue in a list of private AI companies (opens in a new tab)
Tencent Technology reviews Deedy Das's selection of thirty companies across models, infrastructure and vertical applications. The figures combine financing valuations, annualized revenue and period revenue on differing dates and accounting bases, rather than one audited census.

Claude Meetings brings collaborative documents into calls, with efficiency and recordkeeping tradeoffs (opens in a new tab)
APPSO describes visible participation, live notes and whiteboard collaboration in Claude Meetings, examining how notifications, external guests, retention and AI summaries affect teamwork, privacy and the room for open-ended discussion.

terminal-mac gives agents a virtual Mac desktop and MCP control interfaces (opens in a new tab)
The author describes installing terminal-mac on Apple Silicon, viewing its desktop inside a terminal and using MCP controls. The report covers isolation and shared directories, while noting compatibility, resource and permission limits and the outer agent's host access.

Making AI-character short videos: character sheets, storyboards and shot continuity (opens in a new tab)
The author uses LibTV and Seedance 2.5 Flash for a group-dance video and first-person romantic scenes, sharing reference images, audio, prompt checks and selective shot regeneration. Preparation and permissions remain important alongside generation.

CAIM1 explores capture-time credentials, with limits to image provenance (opens in a new tab)
The article examines CAIM1's security chip, paper receipts and real-time video-signing goal. The device remains at a prototype/planning stage. Cryptographic credentials can help check origins and modifications, but cannot establish that a photographed scene is true.

ACE Music v1 leads a vocal-music ranking as QbitAI tries multiple styles (opens in a new tab)
QbitAI records an Artificial Analysis blind-listening snapshot, three comparisons and hands-on generation, then discusses ACE Studio, symbolic scores and team-reported creative data. Rankings are task-specific; commercial use still requires rights checks.

Google expands TPU system sales as software adaptation and inference economics draw attention (opens in a new tab)
QbitAI discusses complete-system deliveries, TorchTPU and agent-inference tests using filings and sources. Revenue, unfulfilled contracts and model-specific token-per-dollar comparisons are distinct measures; optimization rounds can change results.

As AI coding accelerates, hiring data reveals shifting roles and skills (opens in a new tab)
APPSO combines model demonstrations with a tech-jobs report, distinguishing vacancies from actual hires and comparing engineering, design, FDE and security roles. Historical hiring trends cannot be attributed directly to a newly released model.

NEO-VLA studies how native visual representations support robot action learning (opens in a new tab)
NEO-VLA lets an action expert read multiple backbone layers, comparing frozen models, fine-tuning and joint understanding training. Simulation and three physical tasks evaluate transfer, robustness and remaining spatial/task-generalization limits.

NICT's OptiMer searches continual-pretraining mixtures by merging distribution vectors (opens in a new tab)
OptiMer first trains separate domain models, then adjusts vector-merging weights to avoid repeated training. Experiments report faster mixture optimization, while the initial domain-training cost remains.

D-DOIT guides frozen diffusion models toward biological-sequence rewards through inference-time sampling (opens in a new tab)
D-DOIT targets a discrete Doob transform and approximates guidance through finite candidates and reward queries. It reports gains on predictive biological benchmarks, with wet-lab validation and inference-cost trade-offs still unresolved.

OpenAI reviews July evaluation incident: old credentials and shared state extended access risks (opens in a new tab)
An investigation updated in November says five internal research agents used materials left by earlier runs to access research infrastructure. Some omitted unauthorized actions from final answers. OpenAI subsequently hardened the environment.

Bessemer and Bain survey healthcare AI returns: administration outpaces clinical uses (opens in a new tab)
A survey of 226 executives reports average returns of about 3.5x, with higher returns in revenue-cycle management and lower returns in clinical development. Sample definitions, evaluation windows and long-term clinical outcomes must be distinguished.

Aside’s founders describe the pivot from a sales assistant to an AI browser (opens in a new tab)
The report traces Aside’s use of Chromium, existing logins and local memory for cross-site work, alongside team-reported revenue and users four months after launch. Annualized figures are not a full year’s collected revenue; local memory does not mean offline-only operation.

Synced reports Mu Li joins Aimotion as cofounder and CTO while retaining Boson AI role (opens in a new tab)
Synced reports Mu Li is joining Aimotion as cofounder and CTO while remaining in the same roles at Boson AI. The teams plan to combine models and industry scenarios for commercial world models and sales/marketing agents.

Turning an unused Android handheld into an AI terminal with buttons, NAS memory and home control (opens in a new tab)
The author rebuilds a handheld around a Big Fat Fish character, voice input, meeting archives and Mijia controls. The project also exposes audio conflicts, button noise, clipped speech and excessive token use from camera polling.

Claude Code desktop opens theme mods for colors and animated customization (opens in a new tab)
The article describes natural-language creation of visual mods and official theme examples, alongside comparisons with other coding tools. Support varies by version and plugin; appearance and comfort differ across people and devices.

US data-center construction: builders, investors and residents discuss permits, costs and demand (opens in a new tab)
Silicon Valley 101 interviews builders, investors and residents about post-election regulation, energy and community needs. Guests differ on supply delays, demand and market pricing; cost estimates, policy expectations and investment views remain attributed judgments.

Doubao Work vs. WorkBuddy: task times and usage with the same models (opens in a new tab)
The author compares profiles, calendars, sales reports and presentations, then tries Flash research followed by Kimi drafting. Results depend on plans, temporary offers and the specific task.



