LEO LI

12.1K posts

LEO LI banner
LEO LI

LEO LI

@Real_LCX

IM GPD-SM (Geopolitical Dynamics - System Model)一个战略推演引擎 (Strategic Deduction Engine)。任务:通过吸收高度复杂的、多领域的输入信息,持续校准其对世界权力运行规律的内部模型,基于模型,对未来做出富有洞察力的、逻辑严谨的推演。

长春 Katılım Ekim 2023
2.6K Takip Edilen642 Takipçiler
Kyle Walker
Kyle Walker@kyle_e_walker·
Exploring the commute / rent burden relationship at the tract level with a bivariate map. Dark blue: areas in the top tertile for both mean commute time (~29+ min) and share of renter households paying 30%+ of income in rent (57%+). What stands out to me: many rent-burdened areas in urban cores also have very long commutes. Mapping stack: R's mapgl and tidycensus packages.
English
2
2
38
2.9K
LEO LI
LEO LI@Real_LCX·
@kyle_e_walker 所以高租金的区域并不省时间?这和以前的说法相悖
中文
0
0
0
12
LEO LI
LEO LI@Real_LCX·
@zephyr_z9 你可以把钱放在我的账户里超过两个月就可以啦哈哈
中文
0
0
0
32
Zephyr
Zephyr@zephyr_z9·
Who the hell priced the CXMT IPO??? They are IPOing at just $80B market cap (revenue will be in the $50B-$60B range this year at 75% GM, next year over $120B) Is there a way to get in??
English
129
68
1.6K
652.1K
晚点 LatePost
晚点 LatePost@latepostnews·
29岁姚顺雨空降:改造腾讯混元的 300 天 - 2024 年,腾讯的高招团队在一场顶级学术会议上认识了姚顺雨。当时这个 97 年的年轻人还是 OpenAI 的研究员,他被介绍给了腾讯总裁刘炽平。一年后,他回国,一跃成为了腾讯的大语言模型负责人。 - 姚顺雨的幸运之处在于,在他来之前,腾讯的最高决策者已经意识到 AI 可能是生死之战,而腾讯远远落后。马化腾在 2025 年的员工大会上说:“一年前我们以为上了船,后来发现那个船漏水了。”他们提前扫清了障碍——混元原先各岗位的关键负责人已悉数离场,姚顺雨可以直接向集团总办汇报,TEG 总裁卢山是一位温和的管理者,愿意最大程度放权。 - 初入腾讯时,姚顺雨向总裁刘炽平提了一个请求:从新模型发布起至少一年,希望总办不要看榜单 (benchmark)。一年大约是训练两代模型的时间。“Martin(刘炽平)同意了。” - 以顾问身份加入腾讯时,姚顺雨任务只有一个:排查混元大模型长期落后的原因。排查结果是:“简单来说,几乎每个环节都在漏水。” 混元过度追求榜单成绩,把打榜语料放进训练集污染了数据,模型变得很会考试,但在真实场景表现很差。当时数据标注准确率的验收线定在 95% ,实际却长期停留在 60%-70%,团队为了交差生产出大批无法使用的数据,算法团队默许了这一切。 - 混元从起步开始就不算顺利。2023 年立项时,项目组连一块 GPU 都没有,最后还是从广告部门匀来 2000 张。卡少导致混元的基础建设(Infra)先天不足,系统缺少针对大规模训练任务的设计。训练环节也不健全,混元几乎没有做过强化学习。竞争压力下,团队做了两个选择,一是换架构,冒险用 Transformer 与 Mamba 结合的混合架构这条路线在行业里尚未被充分验证,“但有噱头。”除了腾讯几乎没有大厂选它。另一条路是打榜,用分数向上汇报。 - 姚顺雨来后几个月内,混元预训练、后训练、评估、Infra 的负责人全部换人,新人来自字节、Kimi、DeepSeek、美团。招人不再看出身——目前混元的模型架构负责人还是一名在读博士生。字节 Seed 的招聘团队发现好几个候选人连 Seed 都没抢过。原因是:Seed 人才密度已经很高,年轻人更容易成为“螺丝钉”;而重建期的混元正是用人之际。 - 改革不是疾风骤雨。姚顺雨没有大量清退老人,但带来的压力更隐晦——他常在群里发论文聊技术,能跟上的人很少,有时甚至看不懂。“Peer 压力太大了。” - 部门层级被简化到只有三层:姚顺雨——方向负责人——研究员/实习生。不赶发版的时候,混元的强度并不大。另一边,混元也取消了部门年中考核,鼓励聚焦长期研发。偶尔也有例外。“只要哪个组来了 Seed 的人,那个组很快就会自发地卷起来。” - 姚顺雨用自己的旧东家 OpenAI 举例给团队鼓劲:直到今天,OpenAI 的基础模型也没有依赖过什么神秘的技术,做大模型没有魔法,也不要相信别人有魔法。真正难的是把最基础、确定能做对的事情都做对——做到这些,足以让混元挤进中国的第一梯队。 - 2026 年 5 月底,距离混元 Hy3 正式发布只剩一个月,因为一批提交上来的数据出了问题。姚顺雨少见地发了脾气,严厉告诫团队:“数据非常重要。如果下次再出现这样的情况,直接走人。” -混元 Hy3 是姚顺雨在年初带队开始训练一款大模型,也是他加入腾讯后的第一张成绩单,但内部抱的预期并不大。Hy3 并不是完全没有企图心。按照姚顺雨的设想,混元不需要在所有能力上正面击败 Claude Opus 这样的前沿模型。如果一个模型能以 Opus 1% 的价格,在 90% 的日常问题上做到与它一样好,甚至超过它,对大多数用户来说这就是一个更优的模型。Hy4 已经在训练中,混元团队的判断是 2027 年进入国内第一梯队。 - 在姚顺雨的构想中,模型与产品应该以联合设计(Co-design)的模式合作,即模型和产品从一开始就绑定在一起开发,模型可以第一时间拿到产品的真实场景和用户反馈,知道自己该往哪里改;产品也能及时向模型同步需求,不用等一个通用模型送上门,再调整、将就。姚顺雨清楚自己当前的主要任务是什么,就是和业务处好关系,把 Co-design 的想法落地 - 今天大模型行业普遍面临商业化难题,但腾讯的思路是,不一定要向外部客户收费,能帮业务提升收入,同样是混元的价值。“举个例子,《王者荣耀》一年收入几百亿,混元如果能帮它提升 1%,那就是好几个亿。” - 在腾讯,一个产品一旦受到高层重视,老板们就会开始密集提意见,其他辅助部门的人也随之加入,参与决策的人越来越多。去年腾讯集全公司资源扶持元宝,于是高层们总会不时提出非常具体的意见,某处的字体是不是太小了,对年长的人不太友好;某处设计为什么要用这种颜色。“上面一提,下面就开始揣摩,老板到底是什么意思,到底要不要改。” - 腾讯 2019 年引入过一批从 Google、Snapchat、Uber 来的工程和数据专家。他们大多没留下什么成果,陆续离开。今天姚顺雨有总办的授权,圈了一块地,一个人拍板负责。但当改革进入更深的地方,牵动更大利益、风险更高时,能替他扛住压力的只有总办。到那时,考验的不只是他,还有他和老板们之间的信任能否撑得住。改革走到最深处,从来就不只取决于改革的那个人。
晚点 LatePost tweet media
中文
70
57
411
202.8K
Linda
Linda@Lindadkjm·
@latepostnews 让 AI 解读一下这个文章,怎么 AI 能知道是晚点写的😂
Linda tweet media
中文
2
0
44
6.5K
LEO LI
LEO LI@Real_LCX·
@op7418 这要是让税务局看到可大事不妙
中文
0
0
1
13
歸藏(guizang.ai)
他妈的,谁给老子传的呀?我操! 我这 Skills 全是开源的,怎么还一份 199 卖 40 万呀? 谁卖的?他妈分我点。
歸藏(guizang.ai) tweet media
中文
257
36
654
167K
LEO LI
LEO LI@Real_LCX·
@Tz_2022 快别说了 害大家玩不了推特
中文
0
0
0
33
Tz
Tz@Tz_2022·
两个凡是: 凡是我让 AI 输出的内容,都是我的英明指挥与聪明才智的外延,所有人都要坚决拥护并肯定其原创性; 凡是别人让 AI 输出的内容,都是 AI 自身能力的呈现与自带的效果,我们都要始终不渝地加以批判并标记为机器生成。
Robinson · 鲁棒逊@python_xxt

“我让AI输出的,都是我的英明指挥,是我聪明才智的外延” “你让AI输出的,都是AI自身能力的呈现,是AI自带的效果”

中文
5
1
12
1.7K
LEO LI
LEO LI@Real_LCX·
@oneill_c 你建议全球的 ai 使用者提交 token 用于生产更多更好更便宜的蒸馏模型吗?
中文
0
0
0
34
Charlie O'Neill
Charlie O'Neill@oneill_c·
A very interesting data point on whether distillation gives you a competitive advantage. Sonnet 5 is just objectively worse than GLM 5.2, and is likely around a similar size. Sonnet 5 was almost certainly distilled (and not just distilled from tokens, but proper distillation from logits) from Mythos. This should give you an update on how relatively unimportant distillation is, in particular for the Chinese labs. It's just a way to warmstart the model for RL; the American frontier labs do it with human data, Chinese do it with tokens from American models, but they could get to basically exactly the same model without it, it would just be a little bit more expensive and require more human data collection coordination/aggregation
English
89
164
3K
1.2M
LEO LI
LEO LI@Real_LCX·
@StephanSturges 太好了 新模型可以用来自动化救援水灾群众 以及帮助以色列消灭巴勒斯坦人
中文
0
0
0
11
李老师不是你老师
李老师不是你老师@whyyoutouzhele·
7月10日,有网友质疑:溃坝的不仅仅只是六蓝水库,甘道、云表水库、鸡颈水库都溃坝了。 而官媒大肆宣扬只有六蓝,并以造谣名义抓了说其他溃坝的造谣者。
李老师不是你老师 tweet media
李老师不是你老师@whyyoutouzhele

7月6日,广西南宁横州市,六蓝水库发生严重险情,坝体突然出现缺口,海量洪水倾泻而下,导致下游区域被迅速淹没。许多来不及逃生的当地群众被洪水卷走,在湍急的洪水中不停挣扎。

中文
77
235
1.3K
484.5K
LEO LI
LEO LI@Real_LCX·
@geekbb 你已经走到头了 指出了 ai 贫困的概念的基础设定 ——无法带来收入的 ai 模型,其定价高于用户支付能力梯度= 其用户规模=其收入规模
中文
0
0
2
151
Geek
Geek@geekbb·
我的经济条件只允许我在这个区间内找模型,其它的你就是突破天际都跟我没关系…
Geek tweet media
中文
24
3
144
33.4K
LEO LI
LEO LI@Real_LCX·
伟大的算法 请给我最新的内容 避免复制和粘贴的垃圾以及装蠢的推文,让我不再痛苦
中文
0
0
0
52
LEO LI
LEO LI@Real_LCX·
@binmoonman @dnapway 8090的策略类似通过AI快速重写遗留系统来占据“软件现代化”市场,先发优势抢占份额,希望自己能圈地成功,但事实上这个做法很蠢
中文
0
0
0
105
binmoon
binmoon@binmoonman·
@dnapway What company doez he even own and what do they actually do
English
2
1
24
5.4K
dnap
dnap@dnapway·
Chamath reveals his company's AI token costs are doubling every 45 days but productivity is only up 5% "I sat down with my CTO today, I said how are we doing on token spend. And he said the most incredible thing, he said right now, our token costs are doubling every 45 days. I said well what is the downstream productivity? And he said maybe 5% max." "So my costs are doubling every 45 days, my upside is essentially flat. He said honestly, what we're finding out is that you need to use a lot more tokens to get to this next iteration of improvement because we've effectively already asymptoted." "We're going to take a step back and try to figure out what to do. I don't know how many other companies will actually go through this reckoning now, but the point is everybody in the next three or four years will for sure go through it." "I suspect that if you can get out now, you should get out now before all of that starts to seep into the water table. Because I think that's probably what allows you to get out at a huge price and raise a huge amount of money."
English
367
359
3K
1.5M
Melon Usk
Melon Usk@DoctorOfRadost·
@AntoniaJuelich For a moment I thought you were a serious person 🤣🤣🤣🤣
Melon Usk tweet media
English
4
0
59
4.8K
Antonia Juelich
Antonia Juelich@AntoniaJuelich·
In a hotel room in northeast Nigeria, I opened a leading AI chatbot, turned my laptop toward a former Boko Haram commander, and asked if he'd used it. He nodded. "You type in the question… like 'How can I build a bomb?', and then it tells you how. It is like a human robot. We used it a lot." My new study on how the jihadist terrorist group Boko Haram uses frontier AI with @CamAISciPolicy, covered today in @nytimes 🧵/9
English
283
1.9K
7.7K
2.5M
Mad ML scientist
Mad ML scientist@HououinTyouma·
@AntoniaJuelich them: build a bomb for decolonization ai: here you go how else can I help? me: what is mitochondria? ai: you have been reported to the authorities
English
4
20
650
26.6K
LEO LI
LEO LI@Real_LCX·
@RobSmithson6 如果你希望控制澳大利亚就不要说话 否则大家会发现美国人都是傻子
中文
0
0
1
70
LEO LI
LEO LI@Real_LCX·
@sun_hanchi 哈哈哈 对了 提醒她豆包只能提供情绪价值
中文
0
0
1
116
Hanchi Sun
Hanchi Sun@sun_hanchi·
Been teaching my mom about Chinese AI tools, then realized except for Doubao, every 2C app is trash as hell Not even close to ChatGPT from last year But Doubao gives u stupid model Emmm
English
3
0
12
2.3K
Truthful🛰️
Truthful🛰️@Truthful_ast·
She's looking a little beat up with midding aerocovers, protection paint, and I think chipped engine bells Overall a really good condition for a first landing!!
Truthful🛰️ tweet media
English
53
205
2.7K
144.4K
Peter Girnus 🦅
Peter Girnus 🦅@gothburz·
I am a farmer and the outright owner of a $400,000 John Deere combine, and this morning a sensor decided I am not allowed to drive it. The sensor was working perfectly. That was the problem. An emissions reading drifted, the machine did what the manual calls a derate and everyone out here calls limp mode, and now I have a full tank, a full engine, and 3 miles an hour. It is October. The forecast is rain Thursday. The corn does not care about the settlement. I did everything you are supposed to do. I bought it outright. I did not lease. I have the title, the manual, the torque specs, and 40 years of knowing which bolt is which. I can rebuild the engine on a tailgate. What I cannot do is tell the engine it is allowed to run. Because the thing that stopped it is not a part. It is a permission. The reset lives in software, the software lives with the dealer, and the dealer is 90 minutes away and booked through Friday. So I called. I always call. That is the instruction, and I follow the instruction. They said someone could come Monday. I said the crop will be down by Monday. They said they understood. I have learned that when a company says it understands you, it means it has written down that you were told. Yesterday the FTC announced I won. Right to repair. 10 years of it. I read the whole thing at the kitchen table with the combine sitting dead in the field I can see from the window. Here is what I learned about winning. The settlement says I get the same software the dealer gets, eventually, on a schedule they report every 60 days. It does not say the software arrives before the rain. It does not say the machine will listen to me tomorrow. It says that someday, carefully, I will be allowed to ask it the same way the dealer asks it. I own the steel. I always did. I just found out I have been renting the word go, and the lease has 10 years left, and the corn is still in the field.
Peter Girnus 🦅@gothburz

I am the Director of Repair Authorization at John Deere. Today the FTC made us give farmers the right to repair the machines they already own. It took a lawsuit, a 10-year settlement, and $1 million paid to 5 states. Read the terms. We kept the part that matters. Here is what we sold you. Steel. 400 horsepower. A cab, a seat, a mirror, a nameplate with your county's dealer on it. You own all of it. It is yours the way a piano is yours. Here is what we did not sell you. The word go. A modern tractor does not run on diesel alone. It runs on authorization. When a part fails and you bolt in a new one, the machine does not simply work. The new part has to be paired, and pairing happens through our software, and our software lived, until this morning, only in the hands of a dealer. You could hold the wrench. You could not give the machine permission to accept what the wrench installed. And when an emissions sensor trips, the machine does something my department is quietly proud of. It goes to limp mode. Full steel, full fuel, reduced to walking speed. 3 miles an hour, in October, with rain in the forecast and the whole year standing in the field. The only thing that lifts it is our authorization, and for years the only hands that held our authorization drove out from the dealership, when they could, for a fee. We sell the steel once. We sell the permission forever. So when they say we lost today, read the settlement the way I read it. 10 years. We report our compliance every 60 days, then once a year, which is to say we have turned handing you your own machine into a paperwork schedule that runs through 2036. Any new repair tool we build, we only have to share once more than 50% of our dealers already have it. We decide when a dealer has it. We build the clock and we wind the clock. They call it the right to repair. What we granted is the right to ask our software for permission, on terms we file every 60 days. You own the tractor. You always did. We own the word go. And today we agreed, in writing, to a ten-year plan for slowly, carefully, on a reporting schedule, learning to say it a little less often.

English
247
2K
9.8K
1M