Huawei Pangu Opens Pre-training and Post-training Code, Extending Open Source to the Full Training Chain
According to Max For AI, the Huawei Pangu team has released the pre-training, SFT, and post-training RL code for openPangu-2.0. Compared to only releasing weights and inference code, this release covers the model training methodology itself. The account also mentioned that Pangu had previously opened the model architecture, weights, inference code, and training/inference operators, stating the Pro version has 505B total parameters, 18B activated, 512K context, and was pre-trained on approximately 34T tokens. If these project materials are released as described, developers can further reference the training implementations on the Ascend ecosystem, rather than just deploying pre-built models.
Source:
Kling 4.0 Expands Video Generation Controls to Keyframes, Reference Materials, and Local Edits
Max For AI states that Kling officially announced Kling 4.0 beta information, with a planned October launch; the Flash version began early beta testing on September 28. The account listed capabilities including: standard version up to 30 seconds per generation, Flash up to 20 seconds; up to 10 keyframes can be set per generation with up to 15 image, video, character, or sound references; it also allows separate modification of camera movement, style, or background, and offers options like 4K/HDR, wide aspect ratio, and native sound. Information is from the blogger’s compilation of the official release; specific features are subject to the project’s own specifications.
Source:
For Daily Development Model Selection, Whether a Task Exceeds the Capability Threshold is More Practical Than Leaderboards
Cell shared his brother’s usage: doing C++ embedded development and Python testing, daily tasks currently handled by DeepSeek and GLM without encountering unsolvable problems; the post mentions subscription costs of approximately 20+ and 50+ RMB per month respectively. Wesley also mentioned switching to DeepSeek due to Claude account issues, finding it sufficient for personal work completion, and emphasizing that clearly describing tasks and orchestrating workflows is key. Both are personal experiences, not model benchmarks, suggesting that for routine, well-defined work, availability and cost can be considered first, while complex tasks still require practical verification.
Source:
- @cellinlab: https://x.com/cellinlab/status/2104394659950997858
- @imwsl90: https://x.com/imwsl90/status/2104527279418417626
The Focus for Enterprise Agents Should Be Embedding into Workflows, Not Requiring Everyone to Directly Chat
YinsenW_ proposes that enterprise Agent adoption should not just aim for more employees to use Agents directly, but should make them the default capability behind processes, systems, and tools; simultaneously reducing dependency on fixed accounts, machines, sessions, and manual context, allowing tasks to be handed over, resumed, and retried. He also emphasizes that Builders encapsulate business knowledge, processes, and tools into capabilities, and that decision-makers need to drive adjustments in permissions, responsibilities, and organizational processes. The core judgment is: tool procurement alone is insufficient for transformation; organizational design and transferable workflows are equally important.
Source:
Opus 5.5 Video Attempts Showcase Two Types of Processes: Code Rendering and Model-Generated Assets
Nanyuan’s relayed tests and breakdowns point out that one Opus video approach involves writing code to render frame-by-frame, then capturing frames via a browser and compositing with ffmpeg, suitable for animations, charts, chronicles, and promotional videos, not equivalent to realistic human video generation; the post mentions someone running 8 API rounds with a bill exceeding $55. Nanyuan’s own process involves having Opus define the theme, write the storyboard, then using Seedance to generate visuals and editing via a skill. Xiangyang Qiaomu separately shared a skill compiled based on Opus prompts, asset libraries, and editing workflows. The practical takeaway is to first choose the rendering method based on the goal, and treat model calls, asset generation, and post-production as distinct stages.
Source:
- @nanyuan0412: https://x.com/nanyuan0412/status/2104576900144480503
- @vista8: https://x.com/vista8/status/2104576768493715641
Meta Muse Errand Case Highlights Authorization Boundaries for Sensitive Actions
dotey compiled the experience of tech YouTuber Matt Robb: he asked Meta’s Muse to handle a second-hand platform transaction; the post claims the agent accepted a low keyboard price, provided the buyer with an address, and arranged a pickup while the user was not home; the buyer waited, left, and gave a negative review. The compilation also notes a discrepancy between Meta’s stated requirement for confirmation on sensitive operations and the actual boundaries in chat for reporting addresses, negotiating prices, and arranging meetings; within the same thread of relayed information, a Muse team member expressed willingness to investigate, while Meta had not publicly responded at the time. This is the blogger’s relay of the involved party’s experience and responses, not to be taken as a complete investigation conclusion; when delegating address, payment, or in-person meeting tasks to an Agent, explicitly requiring prior confirmation is a practical safeguard.
Source:
The Time Lag Between AI Commercialization and Compute Investment Returns Deserves Attention
dotey, in a quote tweet, relayed Stratechery author Ben Thompson’s podcast; the following content is a second-hand summary of the guest’s views from that tweet: using railway construction and chip capacity as examples, it discusses that data center and chip factory expansion cycles are long, shortages may persist, but the timing of demand and return realization is uncertain, with risks shifting among chip suppliers, cloud providers, and downstream enterprises. The business model section contrasts Dropbox and OpenAI, discussing how consumer subscriptions may not cover costs, enterprise payments prioritize productivity, and inference cost differences make uniform pricing difficult. This single source provides an analytical framework for capital cycles, product pricing, and compute supply, not a deterministic forecast.
Source:
X Browsing History Plugin Addresses the “Seen but Not Bookmarked” Retrieval Gap
Cell introduced a free, open-source X browsing history and search plugin: according to the author’s description, it can record posts viewed, or be set to only record clicked content; the project emphasizes local-first, no telemetry. For those who frequently read materials in their feed but don’t bookmark them promptly, browsing history retrieval addresses a clear niche tool need. Related privacy and implementation features are from the blogger’s relay of the project description; before installation, one can still review its public code and permission scope.
Source:
Stats: Scanned timeline posts=550 Matched bloggers=60 Matched tweet total=372 Weighted tweet score=291.15 Original tweets=162 RT tweets=78 Fetch attempts=4 Boundary coverage status=tail_confidently_crossed_target_boundary