星星之火可以燎原 @fireandstart
(有关必回)(目前是有关注必回关状态)老推友,没有任何目标,混吃等死。 Joined March 2010-
Tweets4K
-
Followers2K
-
Following3K
-
Likes750
这份分析最有用的地方,是把泛泛的“先读博、先学 ROS”拉回到 412 个真实岗位文本:Python 出现在一半工程岗,ROS 只有 7.5%,而操作员和维修岗位也能成为入口。岗位描述不等于最终录用门槛,但这种按公司和技能拆分的数据,比经验之谈更适合用来安排学习顺序。
网店退货规则写得不清,顾客下单会犹豫,商品信息也可能和实际承诺对不上。Google Merchant Center 可把退货政策展示在免费商品信息等位置,但政策必须和官网一致。 设置路径:Products & store → Shipping and returns → Return policies → Add return policy。按国家配置是否接受退货/换货、退货期限、方式、费用和退款处理时间;特殊商品可另建例外,并用 return_policy_label 绑定商品。 先检查官网退货页是否无需登录即可查看,再核对后台与网站的国家、期限、运费/手续费是否一致。Google 会审核政策,未验证前不保证展示。 官方步骤:support.google.com/merchants/answ… 你店里的退货规则,顾客能在下单前一眼找到吗?
CORAL 的关键不只是并行开多个代理,而是把评分器和共享状态接进探索闭环:每个 worktree 可以试不同路线,失败经验也能留下来复用。真正决定它能否扩展到科研任务的,我会看评估信号是否可靠,以及共享知识能否减少重复搜索。
What if you stopped asking one AI agent to solve a problem… …and created an entire AI organization to solve it? That’s CORAL. You give it: A codebase + a grader. CORAL launches multiple autonomous coding agents that work in isolated environments, experiment with different
这类论文清单的价值在于能看到研究脉络如何并行推进:从 JEPA-TTT 的测试时适应,到 World Observer 的潜在世界建模。我会优先追踪它们在长时序、分布外场景里的稳定性,以及能否复现到真实交互任务中。
Must-read papers of the week Top: ▪️ Context Language Models by @Meta ▪️ Invent a Dataset by @adaption_ai ▪️ World Observer by @KAIST_AI ▪️ The Planning Limits of Latent World Models by @uommedia ▪️ JEPA-TTT by @hri_usa and @JohnsHopkins ▪️ False Frontiers by @RutgersU and co
长代理循环里,模型单价不是完整成本:上下文缓存会改变每轮差距,何时升级、带多少有效状态交接同样关键。用真实检查结果触发升级,再只传失败状态和必要文件,才能把预算花在解决问题上。
this is pure f**king gold Anthropic basically hid a routing problem inside a pricing table, and once you model the loop properly, Opus 5.5 stops looking like the “2x model” the trick is that old context gets cheap: > 20K context: Opus is 1.88x Sonnet per turn > 150K context:
把模型按层分到手机和笔记本的 GPU 上协同推理,是个很实在的本地 AI 方向。下一步我最想看端到端延迟、设备间通信开销,以及网络不稳时能否继续生成。
You can now run one AI model across the GPUs of your devices, whatever operating system they run: macOS, Linux, Windows, Android or iOS. In the video, Qwen3.6-35B-A3B, a 21 GB model, runs across a Samsung S25, an iPhone 16e and a MacBook Pro. None of them could load it alone (8
这段演示最有意思的不是“有 6 个机器人”,而是任务能在不同角色间接力:助理拆分日程,旅行代理再协调航班和酒店。要真正省心,还得看交接失败时谁兜底、关键预订是否留给人确认。
SpaceXAI lead engineer (ex-Cursor): "Most people still talk to AI through a chat box. A tiny group already has staff I have 6 bots working for me. At 5am my Chief of Staff hands me the whole day, and while I sleep my travel bot passes the work to a flight bot and a hotel bot"
有播客的人,可能已经有一份不用重传音频的 YouTube 上架入口。关键是:RSS 同步不等于立刻公开,地区资格和最后的发布按钮都别漏。 在 YouTube Studio 里点「创建新播客 → 提交 RSS Feed」,输入 Feed 地址,再验证发到 Feed 邮箱的验证码;选要同步的旧节目范围,保存后等分集上传(可能要几天)。全部上传后,还要去「内容 → 播客」手动点「发布」。 先确认频道所有者所在地区在支持名单里;没有高级功能权限还得做身份验证。适合已有播客 Feed 的创作者,不是把任意音频自动变视频的工具。官方步骤与地区名单:support.google.com/youtube/answer… 你会把旧节目也同步过去,还是只从下一集开始?
这组对比最有说服力的地方,是两种流程用同一份任务说明、通过同一组 24 项检查,再看 Jev 如何接管常规决策,把耗时从 6 小时 24 分降到 1 小时 46 分。自动检查验证控制和计分很有用,但正如后续讨论所说,它不能替代真人试玩;我会把两类证据都放进验收,再根据失败类型决定哪些交给构建者修、哪些需要深入调查。
i gave Dots a 12-second clip and got a playable browser game back four agents handled the build. Jev helped decide what happened next when a check passed, failed or needed another look the brief was one sentence: “rebuild this mechanic as a browser game with keyboard controls,
AI 把视觉稿推进到可运行网站的速度确实惊人,但“像 1 万美元 agency 作品”主要说明呈现力,不等于产品已经可交付。我会把生成后的人工验收放在流程里:移动端布局、键盘可达性、加载性能和真实内容都过一遍;这样提示词负责统一视觉方向,人负责把演示变成可靠体验。
Opus 5.5 absolutely cooked One evening in Claude Code, and the result is a SaaS site that looks like a $10k agency job > glowing light trails across the hero. > a live workflow canvas. > italic serif accents. > one lime accent on a dark theme, held through every section No
我喜欢它把“该不该问人”变成可审计的分类与统计:不可逆、超预算、低置信度、新领域和偏好,分别对应不同的升级理由。再把重复出现的询问转成规则,才能逐步减少打断,同时保留真正需要人工判断的边界。关键是先用真实请求记录验证分类,再调整路由,避免为了少打扰而把高风险决策也自动放行。
a short field manual on how to stop being the switchboard for your own ai agents 27 rules, 4 pages. save it before your next agent asks "should i proceed?"
YouTube 也能做商品联盟,但先别急着算佣金。计划目前只对部分地区的合格创作者开放;能否加入,先看所在地区、是否加入 YPP 和订阅门槛。 申请入口:YouTube Studio → Earn → Programs → Join Now。商品佣金由商家设定,可能变化;退货会冲回佣金,款项通常在购买后 60–120 天经 AdSense for YouTube 结算。 先到 Studio 的 Earn → Shopping 看自己是否有资格和具体商品费率,再决定要不要做。 support.google.com/youtube/answer…
把分类、评分和问答封装成 SQL 函数,确实能缩短数据应用接入模型的路径;这里更值得关注的是批处理把 38 次请求从 23 秒压到 0.86 秒,而不只是“模型进了数据库”。不过项目仍是 pre-alpha,生产选型还要看并发与延迟、失败重试、数据边界,以及模型更新后结果是否稳定。
Jev is now callable from a SQL query. one line. no application layer. that's it. returns a probability. straight from PostgreSQL. the speed that makes it useful: classifying 38 distinct resolution strings from 1,000 NYC 311 complaints row by row: 23 seconds. the same job with
这类对比最该拆开看“单次成本”和“可复现产出”:43M token、52 分钟换来约 70–80% 的效果,确实说明低价模型加长时间迭代可能很划算;但若目标是稳定交付,还得把提示词、人工修订、失败轮次和质量差距一起计入。选模型时,我会用同一任务和验收标准算总成本,而不只看 API 单价。
让 DeepSeek-V4.1-Flash 复刻的这个夏威夷小岛 跑了52分钟,花了43M Token,效果几乎做出来了和GPT-6-Astra 或者 Opus 5.5 的效果相差不大 按 API 的价格计算: Opus - $15.54 Astra - $10.51 DeepSeek - $0.3 相比 Opus 和 Astra 虽然没有完美复刻小岛效果,但达到了70%-80%
这组压缩方法里,我会把“召回”和“排序”分开评估:低精度索引负责快速捞出足够候选,高精度重排改善顺序,但救不回首轮漏掉的相关项。36M 向量低于 30ms 很亮眼;实际选型还要一起看召回率、重排开销和索引之外的元数据占用。PCA/MRL 降维与量化也可以组合,最好用真实查询集测端到端质量与成本。
5 embedding compression techniques, clearly explained: (bookmark this) Ten million 1,536-dimensional embeddings will occupy: - 62 GB in float32 - 15 GB in int8 - 2 GB as packed bits This only covers the raw vector payload, and an in-memory system also needs space for the ANN
文本水印补上了内容溯源的一块拼图,但它回答的是“文本是否带有特定生成信号”,不等于判断作者身份或人类贡献比例。水印强度受文本熵影响,改写也可能削弱信号;因此我会把它和来源记录、检测器误报漏报及独立评测一起看。开放检测器和开源实现,正好能让这些边界接受更多检验。
Announcing our work textGrain, @OpenAI's text watermark in response to the EU AI Act. OpenAI is rolling out textGrain in the EU and opening detector access to researchers. We'll open-source it so the community can build on top.
我最看重 TDD 里的“先证伪”:测试应因缺少目标行为而失败,而不是因为语法或环境问题。这样红灯才能证明用例真的覆盖了需求;实现后再复跑,并保留可重复的检查证据,“修好了”才不是凭感觉。
test-driven-development:先让测试正确地失败,再写最少的代码让它通过。 这是 Jesse Vincent(obra)维护的 Superpowers 里有独立 SKILL.md 的具体 Skill。适合用 AI 写小功能或修 bug
这套流程的价值在于把 Agent 的记忆变成可审查、可恢复的工作状态:先在 worktree 里改,再让 critic 看 diff,最后追加提交,避免直接污染长期知识库。图谱不必一开始就上;当状态跨会话、多 Agent 协作或变更解释真的成为需求时再升级,成本才和问题匹配。
CLAUDE + OBSIDIAN + LOOP ENGINEERING = A VAULT THAT RUNS ITSELF the core idea: the loop keeps its state in the vault, so the chat window can close anytime everything Claude knows lives in a markdown file the loop: > capture - a thought lands in 00-inbox > context - Claude Opus
把 NFT 从铸造时就加密,确实把隐私保护前移了。接下来我最想看的是:合约在看不到明文时如何验证规则,权限如何委托,密钥丢失后资产又如何恢复。FHE 让“可计算而不可见”成为可能,但应用能否建立信任,还取决于这些边界能否讲清并经受独立审查。
GD Afternoon CT ❤️ just applied for @FungoLabs 👀 honestly, the NFT idea here caught my attention. instead of encrypting things later the concept is to have NFTs encrypted from the moment they are minted. with @zama FHE the smart contract can work with encrypted data without
WebCloneBench 把 Agent 操作真实 Web 应用拆成了探索、重建和验证:模型得从有限线索推断界面与行为,再测试自己做出的版本。我更关心分数之外的诊断——哪些应用元素最容易漏、探索步骤如何影响结果,以及跨 Gmail、表格、设计工具等场景的成功率是否一致。这样的评测能把“会操作浏览器”变成可定位、可改进的能力。
this has been one of the most interesting things we've been working on over the past couple of months - it tests the models' ability to autonomously explore a reference site, recreate it, and test what they've built. webclonebench pushes models to understand how web apps work,
Soutine 刘可爱 @Soutine_sayno
4K Followers 4K Following Try Everything! 修炼一个性感大脑「判断力」 终身学习 | 长期主义 让AI成为优质思维的放大器
WAZOBIA CRYPTO @wazobiacryptooo
6K Followers 6K Following ALPHA HUNTER ✧ ONCHAIN WIZARD ✧ ENTHUSIAST
BLANK _ @BLANK53yu
34K Followers 17K Following Blockchain & Crypto Expert | Web3 Enthusiast | PhD in Chart Analysis | Forex expert 30k+ Binance Follower Exploring decentralized finance and innovative tech.
Kausar @Kausar81901420
4K Followers 4K Following Creator & researcher covering DeFi and perps. Threads, breakdowns, campaign work. Open to collabs.
胡子 @Tank300aaa
140 Followers 94 Following
MoonInAI @MoonInAI
6K Followers 612 Following AI Builder |千问办公大使|https://t.co/P056O6becR 折腾过十几种副业的00后,开源自媒体skill全网转发量破万,正在用AI为一百个传统行业赋能(寻找共创伙伴中)
Bubo @bubosees
151 Followers 161 Following I see in the dark . I build AI things and show what's really going on inside.
CC Intelligence Stati... @Changchun_Lens
1K Followers 654 Following Uncover Changchun: Beyond the Ordinary
simran @Aiwithsimran
213 Followers 474 Following AI creator | Tech enthusiast | Turning complex ideas into useful content. Dm for paid collaboration
jiangkoumo @jiangkoumo_
1K Followers 2K Following AI 资讯 | 大模型实测 | Agent 开发 | 终端工具 05大四在读 | 考研备考中 | 学生极客
Tony Nguyễn @tonynguyen_0811
10K Followers 7K Following Co-founder @Vault_Arc Ambassador @sorellalabs Contributor @get_optimum
KK @kellychoi8888
1K Followers 2K Following 看碳市场、ESG披露、绿色债券和 RWA。 拆报告、拆监管、拆谁在承担风险。 不荐币,不带单,不接空投任务。 中文为主,关键帖会补英文。
Wilfredo Sipes @WilfredoS55101
2K Followers 4K Following
lily(互fo) @zhngfng477941
1K Followers 2K Following
• 🐝BuzzFiend-德... @defuliu332748
3K Followers 3K Following 卖家具,也研究加密市场。🪑₿ 一个实体创业者的 Web3 观察笔记:聊趋势、聊生意,也聊走过的弯路。 在现实里做生意,在链上保持好奇。
WangYi @TianAnMen89_64
631 Followers 835 Following 键政,支黑,无业游民,oder,†ʚ pr♡yyds ɞ†,†♡pregabalin♡† 欢迎同类找我聊天,看成分互fo, 新建了一个Telegram闲聊群,欢迎加入一起聊天~https://t.co/oxJvQFvY6L
Jason || Json @jason5819c
1K Followers 883 Following Independent observer. Building hardware by day, decoding society by night. Thinking out loud from the outside. 🇭🇰" 独立观察者。白天造硬件,夜晚解构社会。在喧嚣之外大声思考。
pei @ppeiwoo
862 Followers 807 Following 985硕士毕业转行追逐ai,AI builder,AIPM,自媒体探索者,喜欢拆解产品,喜欢学习前沿ai技术。平时发点碎碎念,也发点个人思考 关注一直限速,等我回关!等我!等我!
名侦探柯零 @Colinclean2
1K Followers 1K Following 有关必回/03数字游民/自学英语口语到母语水平/高强度ADHD式碎嘴/分享有价值的个人跑通的codex项目/互联网精神
Sajid AI @SajidAli_AI
17K Followers 15K Following AI Creator | AI Tools & Tech Promote your brand, product or AI tool here 🎥 Ads & AI Videos for X, YouTube & TikTok 📩 DM for paid promotions & collaborations
dong(互fo) @jzy1008722000
4K Followers 6K Following
禅堂Zendo @14256617zgq
401 Followers 457 Following 🧘讨论AI如何重塑人类心理、关系与意义。 💕在AI时代,重新思考人的连接与存在。 ➡️有关必回,共同创建和谐简中区。
汐神 @jiumeng8852
3K Followers 3K Following 少说话,多观察 🌙 Stay curious 实测跨境收款、虚拟卡和出海工具 AI web3 🇬🇧卡 🇺🇸卡 eSIM 爱好者
NinjaCodex @N1njaCodex
286 Followers 263 Following vibe coding, testing ai models & sharing prompts. building cool stuff without spending a fortune 🥷
Jason || Json @jason5819c
1K Followers 883 Following Independent observer. Building hardware by day, decoding society by night. Thinking out loud from the outside. 🇭🇰" 独立观察者。白天造硬件,夜晚解构社会。在喧嚣之外大声思考。
Sajid AI @SajidAli_AI
17K Followers 15K Following AI Creator | AI Tools & Tech Promote your brand, product or AI tool here 🎥 Ads & AI Videos for X, YouTube & TikTok 📩 DM for paid promotions & collaborations
Young @Young852560
32K Followers 11K Following 女交易员|币圈九年 专注比特币、黄金、美股盘面解析 用 K 线读懂资本博弈,过滤市场消息噪音 TG 交流群:https://t.co/NI7c0Pkqif
YuanshengFu @YuansFu
1K Followers 1K Following ASU ’26 | Thunderbird School of Global Management lost 122lb (55kg) | Maintained for 3 years Share journey| habits | winning mindset Discipline | ENTP
Tajwan Tamim @Tajwan_Tamim
18K Followers 16K Following Content Creator | Web3 AI • NFT • DeFi • Privacy • Gaming DMs open for collabs
Uraz @uraz_yy
10K Followers 9K Following
alucardio @alucardiox
25 Followers 8 Following
Codio @0xCodio
696 Followers 50 Following
名侦探柯零 @Colinclean2
1K Followers 1K Following 有关必回/03数字游民/自学英语口语到母语水平/高强度ADHD式碎嘴/分享有价值的个人跑通的codex项目/互联网精神
Smith @ItsNexoras
2K Followers 2K Following 📚 AI学习|出海探索|钻研投资 🚀 以认知换收益,持续学习,记录成长 🤝 分享所见所学,认识有趣的人,真诚交流
Femke Plantinga @femke_plantinga
14K Followers 610 Following learn with me about AI. growth @SliteHQ
• 🐝BuzzFiend-德... @defuliu332748
3K Followers 3K Following 卖家具,也研究加密市场。🪑₿ 一个实体创业者的 Web3 观察笔记:聊趋势、聊生意,也聊走过的弯路。 在现实里做生意,在链上保持好奇。
MindrAI @Mindr4AI
10 Followers 19 Following
James Briggs @jamescalam
10K Followers 233 Following AI - https://t.co/rLYkOCGxqD - https://t.co/aydfSKDCBB
Asteri @Asteri_eth
3K Followers 635 Following AI writer | Prediction markets researcher | the best vibe and fine content creator, as well as shitposts 🌌
麻酱AI实验室 @zhongying14
6K Followers 628 Following 前大厂产品运营,研究怎么用 AI 赚钱 用 AI 做内容、做产品、做增长、做生意 AI Agent · AIGC · 出海 SaaS · 自媒体变现 合作 / 付费咨询:zhongying97(注明来意)
sopersone @sopersone
6K Followers 222 Following Inexperienced designer | living stress-free | learning quant trading with AI | prediction markets researcher https://t.co/1pLJKcDRcP
禅堂Zendo @14256617zgq
401 Followers 457 Following 🧘讨论AI如何重塑人类心理、关系与意义。 💕在AI时代,重新思考人的连接与存在。 ➡️有关必回,共同创建和谐简中区。
胡子 @Tank300aaa
140 Followers 94 Following
WangYi @TianAnMen89_64
631 Followers 835 Following 键政,支黑,无业游民,oder,†ʚ pr♡yyds ɞ†,†♡pregabalin♡† 欢迎同类找我聊天,看成分互fo, 新建了一个Telegram闲聊群,欢迎加入一起聊天~https://t.co/oxJvQFvY6L
AI Fun @ai_funss
2K Followers 2K Following AI FUN ! ✨分享AI科技 | 🛠️开源工具 | 💡实用技巧 独立开发者的免费开发工具集合: https://t.co/wBo2XnOAwA
Daily Dose of Data Sc... @DailyDoseOfDS_
52K Followers 2 Following Delivering daily insights in DS, ML, RAGs, Agents & AI Engineering. Trusted by over 100k+ readers!
Weijie Su @weijie444
16K Followers 501 Following Training models @OpenAI, Professor @Wharton, CS, Math @Penn, coDir PRiML, PhD @Stanford
KK @kellychoi8888
1K Followers 2K Following 看碳市场、ESG披露、绿色债券和 RWA。 拆报告、拆监管、拆谁在承担风险。 不荐币,不带单,不接空投任务。 中文为主,关键帖会补英文。








































