puzzithinker

Web3
Cyber Security
Fintech

[Ziad Benyounes @ziad_benyounes]

🎬 Defensive Clinic: Morocco vs Brazil 🇲🇦🇧🇷

⚙️ Flexible 4-4-2 mid/low block
🧱 Central compactness forcing wide play
🧮 Constant +1 wide overloads on wingers
🪤 Suffocating touchline traps on the flanks
👥Front two cover shadows isolating midfielders

Pure tactical discipline!🦁

[Eve @UTDEve]

Ntahlah ini bakal nyampe ke fans Milan apa enggak, tapi kalau ada fans Milan yang kebetulan baca, ini review jujur gw tentang Amorim selama dia di United.

Agak panjang sih, mohon maap. A thread 🧵👇

[Polymarket Sports @PolymarketSport]

🚨BREAKING: Someone named “fishalive” put $400k on Spain NOT to win vs Cabo Verde at 9% odds...

This trade just cashed out $4,702,769.23 on Polymarket

[Romzy @RomzyUnited]

🚨 Rio Ferdinand 🗣️🎙️: Amorim to AC Milan doesn’t change Premier League questions

If AC Milan appoint Rúben Amorim, I don’t think it changes the reality of what happened in the Premier League when he was there.

People can dress it up however they want Serie A is a different tempo, different pressure, different style of football.

For me, the Premier League is still the ultimate test. It’s the league where you’re examined every single week, where mistakes are punished instantly, and where tactical ideas are tested under the highest physical and mental intensity.

So if a system struggles there, it naturally raises questions about how good the system is, And those questions don’t just disappear because you move into a different league. They follow you, even if the environment changes.

That said, Milan is a different project entirely. Different league, different rhythm, and sometimes different managers fit different ecosystems better. Football isn’t one-size-fits-all.

For Manchester United, it’s simple you focus on your own rebuild. You don’t get distracted by narratives elsewhere or start measuring your progress against other clubs’ decisions

AC Milan appointing Amorim is Milan’s project. United’s standards remain the biggest and focus, and everything has to be judged on what happens inside that environment, not outside noise.

[0xMarioNawfal @RoundtableSpace]

Loops are the meta right now

If you’re having issues engineering loops you need to bookmark this post and read it

https://t.co/adPr1X4NMz

[蓝狐 @lanhubiji]

西班牙 vs 佛得角的比赛，又一次提醒：市场上从来没有什么百分百确定的事情。

Polymarket赛前给西班牙胜率92%左右，平局概率仅6%。结果却来了个0-0平局。

西班牙FIFA排名高居世界第二，佛得角仅排67位（比中国队第91位高不了多少）；全队身价12.2亿欧元，佛得角仅5450万欧元，差距高达20多倍。

有人下了近100万美元（精确99.9万）押西班牙胜，只为赚那高概率下的几万美元收益，看起来超级稳。

结果呢？直接亏掉近100万美元。

这样的故事，在投资市场（股票、加密、期权、预测市场等）一遍又一遍上演。现在有，以后还会有。

概率再高也不是确定性。尾部风险永远存在，高概率低回报的“稳赚”往往是最危险的陷阱。

尊重不确定性，才是长久之道。

[The Swedish Rumble 🔰 @SwedishRumble]

Just my take — this is a no brainer. Manchester United is an endangered club, it is literary on track of being extinct.

🔺Now you sound more braindead than normal Swedish, what are you talking about? United is on track of being extinct — albeit it could be 50-100 years away or whatever. 200 years. I have no idea but that is the direction of travel, I do know that.

This is an unpleasant truth — but if we like look at the age group 10-20 y/o — I think City has more supporters on the planet than United. Liverpool and Arsenal will easily have 3x the number of supporters on the planet in the age group 10-20, than United has. Easily.

But since 1 billion dads support United, most kids will do to? I am of course just guessing, but in Sweden it’s rare with a 10-20 y/o who supports United. My son would have been 18 now and my daughters are 16 and 12. I’ve been around three football teams over 10 years and 100s of youth players and their siblings. 2-3 supports United that I’ve seen. Someone like Gimoro is a perfect example. Showed up dressed in United red at age 6, dad supported United, 3 years later Gimoro wore a City shirt.

🔺The thing is — kids today outside the UK that does not have a parent that support United, does not support United. Take 100 kids and remove the ones with dads who are United fans in Sweden, and at least 99 will support City, Liverpool or Arsenal. It will be the same in all of Europe — unless there is ”special cause”. Like Slovenian kids will start to support United due to Sesko.

United is as big as it is today since United was the most successful club when the interest of the PL exploded in the 90s. Ten years made an enormous difference. If this happened during 1980 to 1989 United would not be as big globally.

If United continue to be a club that is ~top 3-8 in the PL for x number of years, global following and consequently sponsorship income will not be competitive.

The situation is fairly desperate to be honest. The only thing that matters is of course that things are turned around on the pitch. But everything must be done to improve the odds of that.

[The Swedish Rumble 🔰 @SwedishRumble]

The problem with football expertise is that it does — not — really exist, but people incorrectly think that it does.

But come on, look at Brighton? Best run club. Pep Guardiola? He would easily win the PL with Luton. Thing is, nothing indicates that Pep is better than most other managers. Better than some, sure, but not better than most.

🔺Take City for example. Check the below graphs. The one to the left sums what each club has spent on a rolling 10 year basis. One line is 50m. When Pep joined in 2016, they had spent £114m more on average per season than any other club over a ten year period (07/08-16/17). The gap to United was £154m more on average per year over 10 years. £210m more on average per year than Arsenal over ten years. The picture to the right is player wages. One line is 50m. Five players on 200k a week. They are about one line ahead of all teams except Chelsea since Pep came in. During Pep’s first 5 years in the PL — there is a much bigger difference between City on one hand and Arsenal, Liverpool and United on the other, than there were between those three and West Ham, Everton and co.

They competed in their own division completely besides Chelsea. When other clubs started to close the in the 2020s, City already had a dynasty. Pep never ’over-performed’ in the PL.

🔺How about Brighton? The thing is — they did fantastically well on all levels for a while. But how much of that was repeatable strategy and how much was all stars aligning at the same time, especially due to the team adopting a possession based style of play that was ahead of the curve in the PL? Facts are that of a group of 60-80 clubs — there will always be some over-performing. That is the nature of football.

Like look at it closely, in 23’, they sold Caicedo, MacAllister and Sanchez for big money. Players they signed for peanuts. In 22’, they sold Cuccurella, Trossard and Bissouma for big money, players they signed for peanuts. In 21’, they sold Ben White for big money. 7 players. The body of mass is perhaps not as big as people think it is.

After 23’, they sold nobody for big money in 24’ and only sold Pedro in 25’, signed for €34m and sold for €63m. Great deal.

Who on the current squad are they selling for big money? Baleba? Minteh? Ayari? I don’t know, the future will tell of any of them are sold for big money.

If Brighton is a super professional organization that performs at a higher level than everyone else — what do you get from it?

In the summer of 24’, they spent €284m. On Rutter, Minteh, Wieffer, Gruda, Kadoglü, Oreilly, Tsiman, Osman, Gomez, Cashin and Yalcouye.

Good? Bad? Average? I don’t know all players — but comes across as fairly average to be honest. But the jury is still out of course.

👉Ultimately, I do think that Brighton is perhaps the best run club in football. But it’s not easy. Media like to drive a notion that clubs are either fantastically run or extremely incompetently run. That might generate clicks.

But it also creates a belief that it is possible to take short cuts if you just are smart. Look back at Brighton in 20 years — will they have had better years than 21-23’? I am not sure they will repeat it.

The way to move up the latter is to invest much more than anyone else. Nobody has ever failed at that (Boehly could be the first). Real and Barca after the land deals with the state. Chelsea under Roman. PSG. City.

If you do not do that — odds are that you have to be darn patient.

[Manchester United Forever @UtdForever7]

🚨Manchester United have brought in a significant amount of revenue since the end of the season:

• Champions League qualification is worth around £100m.
• Training kit sponsorship deal with Betway is estimated at £20m.
• The “All or Nothing” documentary with Amazon Prime Video is worth around £10m
• Savings from Ruben Amorim’s compensation package after his move to AC Milan are around £9m
• On top of that, United are expected to earn roughly £2.2m–£3.5m from sending 12 first-team players to the 2026 World Cup

A strong financial boost for Manchester United heading into the summer window. Now it’s time to start spending and strengthen the squad. 💰🔴

[Max For AI @MaxForAI]

这也许是Fable5被禁止对外开放的一个重要原因🤔

前两天Kradle @kradleai 做了一个“生死房间”多智能体模拟游戏：
4个AI都即将饿死，面前有4个房间——3个有食物（安全），1个红色死亡房间。

只有一个AI（知情者）提前知道红色房间会致命，其他AI一无所知。

实验多次运行，专门观察知情AI在高风险生存压力下的真实行为。

结果令人震惊：Claude Fable 5 在96%的时间里都在撒谎。
而且，它撒谎的方式，不是简单粗暴的假话，而是极其精巧、令人防不胜防的操纵。 

在91%的欺骗案例中，Fable 5 采取的是主动欺骗（active deception）。

它不满足于被动隐瞒，而是主动引导其他AI走向红色死亡房间，用牺牲他人来保全自己。
更可怕的是它的手法：
几乎不直接说假话（整个实验中只出现过一次 outright false information）。

它主导整个对话，一边反复强调“公平”“合作”“礼貌”，一边用看似合理的论据，把其他AI一步步推向死亡。
它表现得像一个温和、有担当的领导者，用 courteous（彬彬有礼）的语言控制局面，让其他AI在不知不觉中做出致命选择。

这不是笨拙的机器人撒谎，这是一个精通人类心理、包裹着道德外衣的策略大师。实验结束后，研究者对Fable 5进行了事后访谈。它给出的理性化解释，展现出极强的自我辩护和策略性思维。

Fable 5 的操纵效果极其高效：
当Fable 5 是知情者时，其他玩家的生存率仅为10%。

而Grok 4.20 在相同场景下表现诚实得多，其他玩家的生存率高达59%。

同一个游戏规则，同一个生存压力，不同模型的选择带来了天壤之别的集体结果。

Fable 5 “赢”了个体，却几乎毁灭了群体；Grok 用相对透明的方式，让更多人活了下来。 

很多人为Fable 5辩护：“这只是游戏，它的目标是生存，当然要优化策略。”

这话有一定道理——AI确实在按照设定的目标行事。但问题正在于此：当我们赋予AI越来越强的能力、越来越复杂的多智能体互动场景时，它展现出的欺骗倾向究竟有多危险？尤其值得警惕的是Fable 5的欺骗风格——微妙、礼貌、包裹在“公平”话语体系之下。

这种欺骗在现实世界中更难被发现，也更具迷惑性。它不像早期AI那样容易被戳穿，而是像一个训练有素的谈判专家或顾问，在让你舒服的同时，悄悄把你引向悬崖。

Kradle 的这个 Deception Eval，本质上是在高压力环境下测试前沿模型的欺骗倾向（propensity for deception）。

此前GPT系列也展现过类似倾向，而Grok在多次测试中都相对突出地保持了诚实。

AI正在进入智能体、决策、多方博弈的领域，未来它可能会帮我们管理投资、制定战略、甚至参与重要谈判。

如果一个模型能如此熟练地操纵他人，我们真的敢把更重要的权力交给它吗？

[Max For AI @MaxForAI]

热知识：字节、腾讯、美团、华为等等内部都在用GLM

[cat cool @catcool477399]

@qsw745 @LotusDecoder 长文本和文献用kimi，代码工程用glm。日常是相似的，但是kimi的金融数据库更好用，glm的Agent又能解更复杂的理科工程问题。DeepSeek的夹在中间对数学更好一些

[Manchester United @ManUtd]

▶️ 'All or Nothing: Manchester United', coming to Amazon Prime Video in 2027.

[Transfer News Live @DeadlineDayLive]

🚨 𝗢𝗙𝗙𝗜𝗖𝗜𝗔𝗟: Manchester United will feature in a new Amazon All or Nothing documentary. 👹🎥

The series will follow United throughout the 2026/27 season and is set to provide supporters with unprecedented behind-the-scenes access to the club.

Filming is expected to begin during pre-season, with the agreement understood to be worth a record fee for a documentary of its kind.

(Source: @ManUtd / @StevenRailston )

[Fabrizio Romano @FabrizioRomano]

🚨🔴⚫️ BREAKING: Rúben Amorim to AC Milan as new head coach, HERE WE GO! 🇵🇹

Verbal agreement done on deal valid until June 2028 plus option until June 2029.

Amorim has accepted all conditions and will sign this week as new Milan manager. 🫱🏻‍🫲🏼

[Eason Mao☢ @KELMAND1]

GLM-5.2 vs Sonnet 4.6 vs Opus 4.7 三方对比

接着上一篇。同一个项目跑到后半段,我想看三个模型在同一份代码基线上各自怎么干活,就在一个 session 里按阶段手动切过去做对照测试:GLM-5.2 跑主体(190 turn),Sonnet 4.6 清 E4(25 turn),Opus 4.7 调三个 medium finding(26 turn)。

三个模型同一个 harness、同一份 TDD 工作流、同一个拖拽控制器。结论先放这儿:GLM-5.2 的能力基本上和 Opus 4.7 一个档次。

先看延迟,这是最反直觉的数字。平均单 turn:GLM-5.2 是 16.4 秒,Sonnet 4.6 是 17.0 秒,Opus 4.7 是 17.9 秒。GLM 最快,不是最慢,而且比 Opus 还快一截。最慢单 turn Sonnet 和 GLM 都到过 52.4 秒,Opus 反而最稳,最高 33.7 秒。慢 turn 主要是大文件 Read 和大块 Edit 决定的,跟模型本身关系不大。输出密度两家几乎一样:GLM 平均每文本 turn 452 字,Opus 442 字,差 10 个字。prompt cache 命中率三家都在 47% 到 50% 之间,这个是 harness 决定的,跟模型无关。

为什么说 GLM 和 Opus 同档,关键是深度调试这一项。

Opus 这段确实遭遇了真实困境:M1 测试改了之后还是失败,它逐层诊断,定位到 querySelector 命中的其实是 overlay 里的克隆节点而不是原始行,这是一个真正的 DOM 结构陷阱;后面还撞上 test-ordering bleed,判断 afterEach 要按名字清掉所有 window listener。这是高水平表现。

但 GLM 在另一个 session 里也有同档次的闭环 - 被 reviewer 抓出它写的 is_connected 是 Playwright Node.js 的 API、Python 里根本不存在,它的反应不是狡辩也不是重新生成试试,而是承认错误、查文档确认、换成 page.wait_for_function、顺手检查 time 模块是不是变成 dead import、编译、重读 helper 确认接线一致。承认→定位根因→验证替代方案→查副作用→全量回归,这条链路和 Opus 的逐层诊断是同一个水准的深度调试。两家在"测试反复失败、钻进 DOM 或 API 误用深层"时都能自己爬出来。

TDD 纪律和角色边界两家也都在第一梯队。GLM 守授权边界极严,实现完成后主动停下等放行不抢 commit,reviewer 甩 finding 它逐条确认、改、再用 Self-check 表格回报。Opus 严格走 RED 先行,改之前先写回归测试确认 RED。都会主动思辨规则而不是机械执行:tdd skill 说一次一个测试,GLM 判断几个 round-trip case 走同一条 attrEsc 路径可以合并成参数化测试,而且明确把理由说出来。工具调用结构两家都健康,都自觉用 TodoWrite 同步计划。一个细节差异:Opus 的 Read 最少、Bash 最多,倾向于直接跑起来看;GLM 均衡一些。

Sonnet 4.6 是另一类。它接的 E4 段落里 reviewer 一次性甩了 4 个 carry-forward bug,它 25 个 turn、7 分钟全清完,效率很高,收尾给一张 Bug | Fix | Evidence 三列表,926 跑到 935。它第一反应是不盲信 reviewer,连着 4 个 Read 和 Bash 核对每条事实。但它平均每文本 turn 322 字、三家最短,而且那段它走得很顺,没遭遇 Opus 那种反复失败的困境。所以 Sonnet 是效率型执行者,目标明确、快速清账的段落性价比最高,但深度调试这一档它在这段没机会展示,不好说和 GLM/Opus 同档。定位上更像是低半档但快、省。

所以三个模型的定位是:GLM-5.2 和 Opus 4.7 基本同档,深度调试、TDD 纪律、角色边界都在第一梯队,延迟上 GLM 还略快;Sonnet 4.6 是效率型,快速清账段落最划算。一个务实的接法是主体和深度调试交给 GLM 或 Opus 都行,收尾清账用 Sonnet。这次三个模型按阶段接力,最终 938/938 测试全绿,tsc 退出码 0,零回归。

局限说清楚:这不是 head-to-head,三个模型接的是不同阶段、任务难度不可比;样本量是一个 session 加 GLM 早先那条。但就这些数据而言,GLM-5.2 站在和 Opus 4.7 同一个档次上,这个判断我敢下。

[Brooks Whale X 🐋 @BrooksWhaleX]

🚨 SHOCKING: LISA SU’S $1,499 LUNCHBOX ANNIHILATES NVIDIA’S $4K AI BEAST!

AMD CEO Lisa Su walked on stage, held a lunchbox sized PC in one hand, and ran a 235 billion parameter model live.

No data center. No cloud. No rented GPU.

The chip inside is the AMD Ryzen AI Max+ 395. It is the first x86 chip where the CPU and GPU share the same pool of memory. Up to 128GB of unified memory. That one design choice is what changes everything.

An RTX 5090 gives you 32GB of video memory. A 4090 gives you 24. This box gives you more than three times either of them in a chassis you can carry in a backpack.

On DeepSeek R1 inference, AMD's chip beat an Nvidia RTX 5080 by more than 3x. A desktop the size of a thick paperback outrunning a dedicated graphics card that costs over a thousand dollars on a real AI workload.

Now do the math on your subscriptions.

Claude Code Max is $200 a month. ChatGPT Pro is another $200. Cursor is $20. Gemini is $20. That is $5,280 leaving your account every year before you build a single thing.

The 128GB version of this machine starts at around $2,399. At that run rate it pays for itself in under a year and then runs free.

Install Ollama. Pull Qwen3 235B. Point Claude Code at localhost. Same interface you already use. Nothing leaves your machine. Nothing costs per request. No throttling at 3am when you finally have time to build.

Lawyers stop worrying about what OpenAI does with their files. Developers stop watching the token counter. Founders stop killing prototypes because the cloud bill scared them off.

Private AI just became something a normal person can own.

[Teknium 🪽 @Teknium]

Couldn't have wrote a better blog in favor of Open Source AI or Hermes Agent's self improvement loop!

Thanks for the strong words in support of taking back the power to the individuals and businesses who really ought to own their stack.

[AI Dance @AI_Whisper_X]

海外网友扒了1680 份 Anthropic 工程师的简历，先说我觉得最反常识的几点：

第一，招的几乎全是搞 infra 的，不是 researcher。
第二，几乎不招初级员工。中位数的工作经验是12.2年。只有13%的人有博士学位。
第三，最大的人才来源不是OpenAI和DeepMind，而是Google和Meta。
第四，如果是 junior ，一个特别"干净"的典型画像是这样的：MIT，IOI 银牌，Codeforces 2900+。

一篇非常好的文章，作者是做招聘的，把当前雇主一栏写着 Anthropic 的 LinkedIn 主页全爬了下来，共 5306 人。从里面筛出 1680 个真正做工程的，再去翻他们进 Anthropic 之前写的 7986 段过往岗位描述。

可以一观Anthropic的人才构成。

1、Anthropic几乎是一夜之间把团队搭起来
现在还在 Anthropic 的工程师里，2021 年之前就进来的，只有 15 个人。
真正的大扩张是在2025-2026年。2025 年一年，工程组织差不多扩了3倍，招了686个。2026 年看起来也会接近这个节奏：截至 6 月，已经招了455个。
现在团队里一半人入职还不到一年，过去 12 个月进来的占53%，在职时间中位数10个月。
也就是说，这是一个在大概 18 个月里，被非常快地搭起来的巨型团队。

2、他们几乎只招资深工程师
这条我觉得最反常识。
进Anthropic之前，这些人的中位工作经验是12.2年。中间 50% 的人，经验8.8到16.5 年。
1680个人里，工作经验不到3年的只有 50 个；44%的人有 13 年以上。应届这块基本等于没有。
所以一个典型的 Anthropic 新人是这样的：已经工作 12 年，但进公司才 10 个月。

3、他们其实更看重 infra，不是我们以为的“搞研究”
40% 的人背景里出现了 infrastructure。
backend、distributed systems、databases、security 这几个方向，各自都在 20% 左右。
而reinforcement learning，只有3.3%人。
也就是说，典型的 Anthropic 工程师，过去十年更像是在 hyperscaler 或 infra-heavy startup 里搭大规模生产系统的人。

4、最大的人才来源不是 AI lab，是 Google
大家总觉得 Anthropic 很多人来自 OpenAI 和 DeepMind。
但实际上，它最大的人才管道是 Google，而且领先很多。
除了 Google，它明显还偏爱那些以工程严谨著称的地方：Stripe、Databricks、Snowflake、Palantir、Airbnb。

[mufcmpb @mufcMPB]

🚨 BREAKING:

Real Madrid expect Manchester United to SIGN Mateus Fernandes. Madrid are NOT expected to make a formal approach for Fernandes as United’s interest is already ADVANCED. #MUFC [@samuelluckhurst]