tache

[Wayne @waynoir]

Harvard's CS50 Introduction to AI covers everything from neural networks to the foundations that power modern LLMs, one of the most complete free courses out there

0:00 - neural networks
5:41 - activation functions
7:47 - neural network structure
16:02 - gradient descent
30:00 - multilayer neural networks
32:58 - backpropagation
36:27 - overfitting
38:52 - tensorflow
53:01 - computer vision
58:09 - image convolution
1:08:18 - convolutional neural networks
1:27:03 - recurrent neural networks

Covers graph search, classification, optimization, reinforcement learning and the core ideas behind game engines, handwriting recognition and machine translation

[Mahax @Mahaximus_]

Google Brain founder, Andrew Ng:

"Prompting will die in 6 months.

Loops and Graphs are what's replacing it."

In 2 hours, he shows exactly what the best engineers already build instead, and how to start building it yourself.

The missing piece most people skip: how to connect those loops into a graph that compounds every time it runs.

Watch it, then read the full guide on loops and graphs below.

[AI Will @FinanceYF5]

Palantir卖给政府的战情室系统，一年要几百万美元。

有个叫Elie的人几天内搭了个开源平替，免费放到GitHub上：实时3D地球，接500多个新闻源，AI自动摘要军事、经济、灾害等15个类别，本地跑AI不用API key。

[SC @chenxiaoster]

@FinanceYF5 模型决定单步智能的质量，Agent 系统决定多步执行的可靠性

[RIFF @RIFF404]

省流总结：

好久没看到这么傻逼的发言了，纯正二逼一个，标准的领导发言，而且还是国企里面的那种，看似站在更高维度说话一套一套的，强调观点客观中立，其实完完全全就是经验主义。

他所有的观点我都可以一句话反驳：

1. 对比A社和openai，kimi k3的推理成本非常便宜，但凡真的用过都不会说贵。

2. 无视AGI的风险的通常都是闭源实验室，真正的风险不在开源模型。

3. 如果开源模型直接影响美国的ai资本支出，甚至是闭源实验室的财务，更有可能的原因是你花钱大手大脚，不如多省省，让普通人也用得起AI，而非权贵主义产物。

4. 了解下什么是社会主义，什么是共产主义，多读读书白左精英们，开源权重模型不属于任何一个类别。

5. 那为什么是Anthropic首先在模型添加后门呢，到底是开源风险大还是闭源？可能政府还没ban，模型先把用户ban了。

6. 这个问题属于扩散焦虑，可以散了。

总结：目前看有理由怀疑这个人跟深层政府/deep state打过交道，充当发言人的角色引导公众舆论和美国头部ai实验室的战略方向，从创新合作至怀疑对抗。纵观历史，任何新技术的发展和革命都在打破普通人的信息差，为人民服务。在这个时代，旧权贵们除了钱和权，在认知和信息端已经不及一个开源模型，你说谁更慌呢？

[Archive @ArchiveExplorer]

My friend at Anthropic makes $1.4M a year.

No MIT. No Stanford. No PhD.

I asked how their team actually ships autonomous agents.

He sent me an article most AI engineers missed.

It unpacks Anthropic's entire loop-engineering playbook: the 3-agent harness, all the team's internal essays, and how to copy it.

https://t.co/kNNDjtD1Y2

Bookmark it. Then read this setup ↓

[Bob Fu 傅丰元 @fm100]

读 @thinkymachines 的最新技术路线博客时，发现他们把汉娜·阿伦特的《人的境况》、迈克尔·波兰尼的「隐性知识」，以及哈耶克关于「分散知识」的讨论，都放进了引用文献里。

Thinking Machines Lab 是 OpenAI 前 CTO Mira Murati 离职后创办的 AI 公司。它在成立几个月后便完成了约 20 亿美元融资，投后估值达到 120 亿美元。

最近，他们发布了一篇博客 《The Future Worth Building Is Human》，介绍自己的技术方向：不只追求更强的通用模型，也希望个人和组织能够继续训练、定制自己的模型，让人的经验、知识和判断参与 AI 的工作过程。

很有意思的是，这样一家站在前沿的 AI 模型公司，在解释技术路线时，引用了不少哲学、经济学和社会思想。

1966 年，迈克尔·波兰尼《隐性维度》

波兰尼提出了「隐性知识」：我们知道的，往往比能够明确说出来的更多。

厨师对火候的判断、工程师排查问题时的直觉、一个团队长期形成的工作默契，都很难完整写进文档或数据库。Thinking Machines 用它来解释，真正有价值的知识常常产生于具体实践，也会随着实践不断变化。

1945 年，哈耶克《知识在社会中的运用》

哈耶克讨论的是「分散知识」：社会中的知识并不集中在某一个地方，而是散落在无数个人、组织和具体场景里。

一家餐厅如何调整菜单，一间商店如何摆放商品，不同组织如何完成同一件事，都依赖身处其中的人掌握的局部知识。Thinking Machines 借此讨论，单一的中央模型很难吸收所有持续变化的经验。

1955 年，约翰·冯·诺依曼《我们能在技术中生存下来吗？》

冯·诺依曼认为，技术中有益和有害的部分总是靠得很近，很难把「狮子」和「羔羊」彻底分开。

Thinking Machines 用这段话讨论模型的开放、所有权和安全：让人拥有更深度修改模型的能力，也意味着这种能力可能被用于不同的目的。

1958 年，汉娜·阿伦特《人的境况》

文章最后引用了阿伦特关于「手段与目的」的讨论。

当一切都只按照效用来衡量时，人也可能变成实现某个目标的工具。只有当人本身被视为目的，所谓的效率、工具和技术，才重新获得意义。

原文：
https://t.co/5FIe8Na8Yt

[indigo @indigox]

AI 实验室里最懂行的人都在说同一句话：“一年后人类就没什么可做的了。” 你的 Harness 不重要，AI 会自己生成；你甚至不再是终端用户，因为它在跟别的 AI 说话；材料、物理、数学、健康 …… 根本性难题将被逐个接管！来自 Naval 最新 Podcast「Live in the Future」👀

未来只剩两种职业，要不在 Anthropic 上班，要不为 Anthropic 员工提供（性）服务；剩下的人，请在面包救济队里排号😆

你不是独自生活的，你活在一群轰鸣奔涌的人群里；当这群人决定冲向悬崖，你很难独善其身。现实已经在发生：有人开始攻击数据中心，理由说是「耗水」，但那是个被揭穿的借口。愤怒需要一个出口，没有水，就会有别的。

今天和朋友聊天正好说到，未来只有 AI 巨头和苟且为生的小公司，大家正在做的工作和一些小生意，最大的意义就是让自己快乐和满足、以及糊口，但“糊口”这件事情政府应该会来解决！

所以，现在开始要找一个让自己快乐的差事，然后把剩下的钱全部拿来买 AI 巨头的股票，这是除了在 Anthropic 上班和为Anthropic 员工提供服务的第三种选择😜

[Yero @Yerocode0]

Google acaba de soltar un curso de solo 1 hora sobre Ingeniería Agentiva que destroza a la mayoría de cursos de pago 🔥

Timestamps:
00:00 → Cómo construir tu primer agente de IA
08:24 → Memoria de agente (corta, persistente y larga)
28:34 → Bucles agentivos y agentes de larga duración
40:04 → MCP vs API (esto solo ya vale la pena)
1:00:22 → Sistemas multiagentivos

En 60 minutos aprendes más que en 10 cursos pagos. Míralo hoy. Luego lee el artículo y construye un sistema agentivo que se auto-mejora solo.

¿Lo vas a ver ahora o lo guardas para “después”? 👀

[h100envy @h100envy]

Alibaba engineer who leads Qwen explained the future of open agent models in 25 minutes - better than $2000 LLM training courses.

pre-train the base ->SFT -> RLHF -> tool use -> multi-modal -> ship a whole family (chat / VL / coder / math / QwQ).

That loop is why Qwen quietly became the most downloaded open model family on Hugging Face.

Qwen base + Qwen-VL + Qwen-Coder + QwQ reasoning - that's the stack.

Watch and save it, then read the article below.

[Amber (Jiachen) Liu 🍊 @JIACHENLIU8]

https://t.co/9gMes8x7sa

[Akshay 🚀 @akshay_pachaar]

https://t.co/q4DvRIfStN

[Jason Weston @jaseweston]

Claim: Autoresearch that moves the frontier will be about better data: we call that *Autodata*.

🧵1/6 -- Paper is out! https://t.co/b8gOALndzy

Key idea: agentic data creation provides a way to *convert increased inference compute into higher quality model training*.

We show our method gives gains on computer science, legal and math problems over classical synthetic dataset creation methods.

We also show how to train (meta-optimize) such a data scientist agent, so that it can create even stronger data.

Overall, we believe this direction has the potential to change how we build AI data!

[Ruuj @RuujSs]

An algorithm that turns $1 into $36 billion over 22 years sounds like a false headline right.

It's actually Table 3 of a University of Johannesburg working paper (Nkomo & Kabundi, ERSA 394), a Kalman-filtered, momentum-extended Anticor algorithm (K-ACM), backtested on NYSE data from 1962–1984.

No capacity constraints. 10bps in costs, full liquidity assumed.

The mechanism trade on deviation from Kalman trend, not raw returns is genuinely interesting. The headline number is a backtest artifact.

Bookmark this!

[h100envy @h100envy]

Here is a condensed, high-impact version of the post:

Discover -> Hand off -> Verify -> Persist -> Schedule

This 12-page PDF completely changed how I build agentic systems. Here is the 5-step blueprint:

Discovery: The loop reads CI, issues, and commits to find what's worth fixing.

Handoff: Each finding gets an isolated git worktree so parallel agents never collide.

Verification: A second agent - built to assume the code is broken, reviews the work.

Persistence: Results land on disk, never in a temporary context window.

Scheduling: Automation fires the entire process on a timer, making it a true loop.

The key insight: An agent grading its own work always praises it. You need that second agent to say "no."

Read it now, then explore the article below.

[monokern @monokern]

https://t.co/RmlXgqYeJg

[Roan @RohOnChain]

https://t.co/JRsWSeUfrN

[tache @tacheme]

https://t.co/EN8RKXEXlZ

[Roupen Odabashian @RoupenMD]

The most important AI-in-medicine result of the last two years is the one almost nobody wants to repeat. Giving doctors a frontier model did not make them better diagnosticians.

In a randomized trial, physicians using GPT-4 plus standard resources scored 76% on diagnostic reasoning. The control group, using UpToDate and Google, scored 74%. Statistically a tie.

Here's the twist that should keep us up at night. The model alone, with no doctor attached, scored about 16 points higher than the physicians it was supposed to be helping.

So the AI was excellent. The doctors were fine. And the combination added almost nothing. The bottleneck wasn't the model's intelligence. It was the human-AI interface. We glance at the output, anchor on our first impression, and use the tool to confirm rather than to think.

That reframes the whole problem. We've spent two years racing to build models smart enough for medicine. This says the harder, less glamorous work is teaching clinicians how to actually use one. Capability was never the gap. Adoption behavior is.

What would it take to train a doctor who gets more than 2 points out of a system that's already better than them alone?

[alphaXiv @askalphaxiv]

“Do as I Do: Dexterous Manipulation Data from Everyday Human Videos”

With how robot dexterity is bottlenecked by data as teleoperation and MoCap are expensive and internet videos are only observational, this paper turns normal RGB human videos into executable robot hand trajectories.

It reconstructs the hand and object, tracks the object with SAM 3D guided diffusion, then retargets the motion through physics-aware optimization so a real dexterous hand can do the same task.

The result is a path from internet video to real robot rollouts, with retargeting success improving from 25% to 71% on noisy reconstructions.