Web3 老尘
Web3 老尘
2017年入圈，经历了各种起起伏伏活到了现在，不认命，但认输
5Following
72followers
Feed
Feed
Someone used dsh-blender-plugin to independently complete a Blender modeling and rendering job. The bill was 1.58 eTokens, taking 1 hour and 20 minutes, equivalent to about 12 RMB. When 3D outsourcing quotes and delivery times encounter this cost scale, they all need to be recalculated. The plugin is already publicly available on GitHub; if you want to try it, just download and run it once to see how deep it goes.
Chimney
Shocked and stunned!!
I used this plugin inside to let the ds1017 independently complete the modeling and rendering effect.
Cost: 1.58eToken 1h20min 12rmb
Below are the effect images and other works modeled by group friends using 1017～
Everyone is welcome to download dshTUI and dsh-blender-plugin to experience it!
Looking forward to seeing everyone's AI-generated modeling works 👏
Tencent Hunyuan has upgraded the intelligent agent memory framework Hy Memory to version 2.0 and fully open-sourced it: a three-layer structure including a dedicated model, memory framework, and MemHub application, each layer can be used independently or combined into a complete solution.
Regarding memory, the ability to "remember you" during AI chat, "remember project context" during work, and "remember what was changed" when coding are three distinct capabilities. The official evaluation separates these into chat, work, and coding scenarios, and it ranked first in open source in all three.
It is already running in Tencent's self-selected stocks and Tencent News business.
I noticed that today the Chinese-speaking community is buzzing about Claude account bans.
Anthropic itself disclosed that it banned 11.4 million accounts in the first half of the year, with only 42,000 appeals, a negligible ratio. Some accounts were banned just 15 minutes after registration, without sending a single message; the appeal portal is completely blocked, and tickets are automatically closed by bots. On the 12th of next month, a new batch of terms will take effect, and the community has already started discussing the next round.
What is most alarming about this is not the bans themselves, but that many people's workflows, prompt templates, and Agent configurations are all stored within their accounts. Once an account is banned, all the work accumulated over two years is wiped out.
Only today did I clearly realize:
Thinking deeply and being able to explain clearly on the spot are truly two different skills.
The more stuff in your head, sometimes the easier it is to get your answers mixed up.
From now on, when suddenly asked a question, I want to practice one small thing first:
Don't rush to prove you've thought a lot.
First, answer the exact question the other person is really asking.
OpenCode, this open-source programming agent, directly integrates free models, allowing it to run without logging in. If you want to upgrade, you can connect your own Claude or GPT. It has over 200,000 stars on GitHub and 16 million monthly active developers, focusing on not storing code or context. Previously, such tools were always tied to a paid account, but now with "built-in free models," the barrier is reduced to zero.
Try it on local projects first, and then decide whether to connect to paid models; in privacy-sensitive environments, it offers one less concern compared to cloud-based closed-source tools.
Chainalysis's report says the U.S. has become a net importer of Bitcoin again, the last time being in 2022. Breaking it down shows a split: small amounts under $10,000 have been flowing out continuously, while large amounts over $10 million have seen a net inflow of about 53,000 coins, or $4.3 billion, in one year. Retail investors are selling, institutions are hoarding; the U.S. treats crypto as reserves, while other places use stablecoins as payment tools. This data reflects behavioral differences, not market predictions. I don't recommend using it as a basis for trading, but it illustrates one thing: on-chain capital flows reflect who is actually making moves better than sentiment does.
I used to think that when working on a long-term project with AI, it was best to stay in the same conversation the whole time.
But once the project extends, it gets mixed with many discarded ideas, old versions, and temporary decisions.
So now I prefer to break long-term projects apart.
When a phase ends, I keep only what really needs to be continued:
Where we are now,
Which decisions have been finalized,
Which plans have been abandoned,
What the next phase is preparing to do.
Then start a new conversation to continue.
SOL's Alpenglow upgrade plan activates on October 21.
Final confirmation time target: reduced from 12.8 seconds to about 150 milliseconds.
Validator fees: fixed 0.8 SOL per epoch, directly burned.
However, these two figures alone do not prove a synchronous increase in SOL's economic value.
After mainnet launch, focus will be on comparing actual net burn amount, number of validators, and real user transaction volume.
DeepSeek V4.1 Flash entered AA-AgentPerf: 552B total parameters but only 16B activated. Officially, it requires at least 2 GPUs to run, with an intelligence score of 39, surpassing V4 Pro's 36. On the leaderboard, GB300 ranks first with 261 million tok/s/MW per watt. The old rule of thumb for local deployment—bigger total parameters mean better model—no longer applies. You have to look at activated parameters and throughput per watt to know which runs more for the same electricity cost.
