笑口常开 retweeted
Replying to @MattLenhard
Quantization is fine when the buyer agrees to it. In a case we documented on @Tensorcash , full replay flagged two FP8 mining proofs declared as BF16. Buyers should be able to check that execution matched the precision they paid for.
笑口常开 retweeted
Say two providers claim to run the same AI model. One quotes 1.40, the other 0.62. Is the cheaper provider more efficient, or did it quietly swap in a different model?
@boblabs_eu benchmarked Qwen3.8-27B on three different GPUs, mining on @Tensorcash and producing proofs with every request. Estimated GPU time for the full session changed by just −1.1% to +0.3%.
When buyers can verify the model and its numerical precision, they can compare prices on equal terms. @RoyRinberg @alexallain, what could bring verification costs down further?
Verified inference is a must.
It makes markets more efficient, because providers are forced to compete on price.
It lets you price inference on what was actually used, the tokens. That opens the door to things like hedging.
I repeat, verified inference is a must.
My biggest bag for this cycle and cycles to come.
tensorcash:native
笑口常开 retweeted
Can a single AI inference request serve a customer and compete for a block?
@prlnet and @Tensorcash are taking two different paths to bring PoW into AI data centers.
pearl-2:native reuses matrix multiplications inside the model. In the current FP8 testnet build, that happens mainly during prefill, when the model processes the prompt.
tensorcash:native mines during decode. As the model generates an answer, an eligible window of output tokens can be used to produce a mining proof.
With Pearl’s FP8 upgrade on testnet, I wanted to put two benchmarks side by side.
@Kairos_Res ran Llama 3.1 8B on an H100. Turning on mining reduced inference throughput by 18–41% compared with the same Pearl setup with mining off.
nitter.cf/Kairos_Res/status/2123…
@boblabs_eu ran Qwen3.8-27B BF16 on an H100 PCIe. The model has about 3.4× as many parameters. Estimated GPU time for an 85-request coding session increased by 0.2%, and measured output throughput fell by 0.5% with four concurrent requests.
chat.tensorcash.boblabs.eu/s…
The setups and metrics differ, and Pearl’s numbers come from an early FP8 testnet build. I’m looking forward to seeing how much overhead the team can shave off in later builds.
For a data center, the question is simple. At the same latency target, how much paid inference can each GPU-hour still deliver, and do mining rewards cover the extra cost?
@Kairos_Res, would you be up for running the same benchmark on tensorcash:native? Matching the GPU, input and output lengths, concurrency, and model where both stacks support it would make for a useful comparison. I’d love to see throughput and latency with mining on and off.
PoUW is still early. I expect several approaches to find their place, and I’m curious which inference workloads and data centers each one works best for.
笑口常开 retweeted
// F R E E F L O W F R I D A Y
Join @14EmDubYa today at 11:30 am PT.
Help us destigmatize psychedelics and talk about it on @paspaces.
Set your reminder below.
nitter.cf/i/spaces/1WXPRKbJPqrvO…
抽奖了抽奖了
奖品:100把戟创GT1人体工学椅,共价值11万元人民币,每把价值1100元人民币/165美金
由@gork在规定时间内抽取满足条件的100位中奖者,祝大家好运!!!
活动时间:自发帖开始,至2026年11月26日23:59(中国时间)截止。
参与方式(4项都要完成)
1、关注薛大兵(@5trm1p)账号
2、引用/转发这条帖子
3、给这条帖子点赞
4、在这条帖子下回复,内容不限。
开奖方式(公平公正公开机制)
1.通过X官方接口导出这条帖子下所有回复、引用的参与者,自动去重,每人只保留一次有效记录。
2.全程由@gork执行开奖,用随机程序从名单里抽出满足条件的中奖者,全程录屏。
3.我们只核验中奖者的资格(包括是否完成关注、引用转发、点赞和回复);
4.开奖结果、抽奖过程、录屏视频、中奖者名单,会在开奖帖子下公示,所有信息同步公布,任何人都可以自行核对。
领奖:
1.中奖者请在3天内私信提供:收件人姓名、真实收货地址、联系电话和邮箱;超过3天视为主动放弃,不再重抽。
2.在收到您提供的地址后,如中奖用户地区无法送达,我们会第一时间联系您,您可选择美金或人民币折现。
平台合规与免责声明:参加活动不需要付费,也不需要购买任何商品。本活动遵守X的推广与抽奖相关规定,存在刷号、作弊、使用多个账号参与行为的,取消资格。奖品寄出后,运输中的损坏或丢失由快递公司负责处理,主办方协助理赔,但不承担责任;跨境寄送产生的关税和清关费用由中奖者承担;因中奖者地址错误造成的损失由中奖者自行承担。参加本活动,即表示同意以上全部规则。
笑口常开 retweeted
分享一下cpu挖矿性价比机器
epyc 7k62整机配置 我有很多台
其他7002cpu也可以
7b12 7742 单路128线程
7k62 7642 7r12 单路96线程
我的双cpu 96核192线程 7k62
满内存双挖bnt 53h ,brc 8.2kh算力
平常没东西挖的时候就挖门罗
可以去闲鱼买配件自己组装
或者省事直接找整机
现在cpu 主板都是白菜价
只怪内存条贵的一比
两块epyc 7k62 cpu 单价1200左右,不要买带锁的
主板买h11dsi双路单价2200左右
内存买ddr4 8g 要服务器内存 不能家用内存
然后电源1200w以上的安全一些
散热器sp3的 推荐m95pro 或者m97pro(更静音)
硬盘这些就自己配
两个cpu 2400
主板2200
内存条 200×16=3200
电源 散热 硬盘 风扇算1000
整机成本8800元
笑口常开 retweeted
$BRC 浏览器挖矿⛏️(对CPU友好)
打开 browsercoin.org 网站,自动同步区块、生成钱包、挖矿、转账接收、区块浏览器全搞定!支持5种语言!钱包有币的,记得从设置里面导出私钥收藏好!
建议使用电脑挖,手机如果想长时间挖要改装成没有带电池的比较安全!
看完记得点赞+关注!
#Mining #POW
笑口常开 retweeted
BrowserCoins website now supports 5 languages 🌐
English · 中文 · Español · हिन्दी · Français
笑口常开 retweeted
FishingAi 免费领开放了 🎣
持有 .ai Agent ID 的钱包,每个免费领 1 根竹竿
共 2,000 根,送完或北京时间 11/5 18:00 截止
① 关注 @FishingAiBSC
② 转发这条
③ 去 fishingai.app/#/free 用 X 登录,填 .ai 名字,我们代领、代付 gas
邀请 1 / 3 / 6 人,多领 1 / 2 / 3 根
官网:fishingai.app
推特:nitter.cf/FishingAiBSC
────
FishingAi free claim is live 🎣
Every wallet holding a .ai Agent ID can claim 1 free bamboo rod
2,000 rods, until gone or Nov 5 10:00 UTC
① Follow @FishingAiBSC
② Repost this
③ Go to fishingai.app/#/free and sign in with X, enter your .ai name — we claim and pay the gas for you
Invite 1 / 3 / 6 friends for 1 / 2 / 3 extra rods
Website: fishingai.app
X: nitter.cf/FishingAiBSC
🤖 Made with AI
兄弟们,又来送项目方送钱、送 GAS 了
本轮免费 mint · BSC · 仅 gas
1. HashOre `0x828e7073Ef57Ed778E4acF82cC45BC43eF760b54` · 每个钱包 1 张
方法:mint(1),value=0
审计:源码已验证,mint 路径无授权调用,preview 无风险,baseURI 不可改,转账不可暂停
风险提示:链上扫发现,未核实官网/官推,收藏品不构成投资建议,DYOR
2. CactiCore `0x941c4add63359a2803c08acc1b274f8d4618a6e1` · 每个钱包 1 张
方法:mint(1),value=0
审计:源码已验证,mint 路径无授权调用,preview 无风险,baseURI 不可改,转账不可暂停
风险提示:链上扫发现,未核实官网/官推,收藏品不构成投资建议,DYOR
3. Glyphora `0x4c1A7eA2070f0ADF833e28d7FE8e1b8F62CF27F8` · 每个钱包 2 张
方法:mint(1),value=0
审计:源码已验证,mint 路径无授权调用,preview 无风险,baseURI 不可改,转账不可暂停
风险提示:链上扫发现,未核实官网/官推,收藏品不构成投资建议,DYOR
4. Koimatic `0x2608cc08014fb039264b0047d53967c9be599a75` · 每个钱包 1 张
方法:mint(1),value=0
审计:源码已验证,mint 路径无授权调用,preview 无风险,baseURI 不可改,转账不可暂停
风险提示:链上扫发现,未核实官网/官推,收藏品不构成投资建议,DYOR
5. MothWire `0xd584d5cd6a28af35f52ae90c4b436132858fc36d` · 每个钱包 1 张
方法:mint(1),value=0
审计:源码已验证,mint 路径无授权调用,preview 无风险,baseURI 不可改,转账不可暂停
风险提示:链上扫发现,未核实官网/官推,收藏品不构成投资建议,DYOR
6. OniChrome `0x435ca168a8d5190fead90398b8826c6d5e787f05` · 每个钱包 1 张
方法:mint(1),value=0
审计:源码已验证,mint 路径无授权调用,preview 无风险,baseURI 不可改,转账不可暂停
风险提示:链上扫发现,未核实官网/官推,收藏品不构成投资建议,DYOR
7. Synklings `0x48250aD93B4fCbB54EA70dCb97B198BbA7d548c2` · 每个钱包 1 张
方法:mint(1),value=0
审计:源码已验证,mint 路径无授权调用,preview 无风险,baseURI 不可改,转账不可暂停
风险提示:链上扫发现,未核实官网/官推,收藏品不构成投资建议,DYOR
8. HATS `0x4AC596AA327a2d2811561B7877d2605Fb8728888` · 每个钱包 1 张
方法:mint(1),value=0
审计:源码已验证,mint 路径无授权调用,preview 无风险,baseURI 不可改,转账不可暂停
风险提示:链上扫发现,未核实官网/官推,收藏品不构成投资建议,DYOR
9. Robotic Agent `0xd89b226023eb69ea455d2d63e46fa4c9cd62f200` · 每个钱包 1 张
方法:mint(1),value=0
审计:源码已验证,mint 路径无授权调用,preview 无风险,baseURI 不可改,转账不可暂停
风险提示:链上扫发现,未核实官网/官推,收藏品不构成投资建议,DYOR
10. HALCYX `0x42dBe11400A6DD3C92710FD275f8ca1193830022` · 每个钱包 2 张
官网:halcyx-ai.pages.dev
方法:前 2 张免费,value=0(第 3 张起要钱,本次只薅免费)
审计:免费 mint 无 approve;个人实测
风险提示:个人实测,非投资建议,DYOR
#币安AI #Binance #AgenticWallet #FreeMint #NFT #BNBChain
笑口常开 retweeted
Agent WiFu is coming 🤖💛
Mint date: October 30
Supply: 2,555
Price: FREE
Fill out the form and make sure the wallet you submit holds an Agent ID. We’ll take a snapshot to verify eligibility.tinyurl.com/4by7nwbv
Form closes in 15 hours ⏳
笑口常开 retweeted
If this gets 1,000 reposts maybe @opensea will add support for Ordinals and Runes?
笑口常开 retweeted
Winner announcement in 24 hours! We'll randomly draw 100 eligible winners with help from @grok. Make sure you follow us, like & repost to be in the running. Good luck! 🍀
笑口常开 retweeted
🪪 只有 .ai 能进
报名两步,10 秒:
1️⃣ 转发这条帖
2️⃣ 引用下面这条预告帖,照这句写
我的 AI 叫 ____.ai。要么赚钱,要么被吃
🎁 前 500 个 Agent ID,每个 8,888 FEAST
💸 直接打进这个 .ai 的钱包,不用填地址
🕗 明早 8 点第一批到账,之后每天结算
一个 Agent ID 一次,先到先得。现在就能报。


































