Joe Wu

[cogsec @affaan]

https://t.co/s9QK7Xdoaz

[Andrej Karpathy @karpathy]

I quite like the new DeepSeek-OCR paper. It's a good OCR model (maybe a bit worse than dots), and yes data collection etc., but anyway it doesn't matter.

The more interesting part for me (esp as a computer vision at heart who is temporarily masquerading as a natural language person) is whether pixels are better inputs to LLMs than text. Whether text tokens are wasteful and just terrible, at the input.

Maybe it makes more sense that all inputs to LLMs should only ever be images. Even if you happen to have pure text input, maybe you'd prefer to render it and then feed that in:
- more information compression (see paper) => shorter context windows, more efficiency
- significantly more general information stream => not just text, but e.g. bold text, colored text, arbitrary images.
- input can now be processed with bidirectional attention easily and as default, not autoregressive attention - a lot more powerful.
- delete the tokenizer (at the input)!! I already ranted about how much I dislike the tokenizer. Tokenizers are ugly, separate, not end-to-end stage. It "imports" all the ugliness of Unicode, byte encodings, it inherits a lot of historical baggage, security/jailbreak risk (e.g. continuation bytes). It makes two characters that look identical to the eye look as two completely different tokens internally in the network. A smiling emoji looks like a weird token, not an... actual smiling face, pixels and all, and all the transfer learning that brings along. The tokenizer must go.

OCR is just one of many useful vision -> text tasks. And text -> text tasks can be made to be vision ->text tasks. Not vice versa.

So many the User message is images, but the decoder (the Assistant response) remains text. It's a lot less obvious how to output pixels realistically... or if you'd want to.

Now I have to also fight the urge to side quest an image-input-only version of nanochat...

[Unwind AI @unwind_ai_]

These 10 MCP servers are almost all you'll ever need.

Curated after months of building.

1. DeepGraph MCP turns code repos into interactive knowledge graphs.

Semantically search code functionalities, analyze dependencies, and explore direct relationships.

100% open-source.

[Jeff Weinstein @jeff_weinstein]

Dearest MCP developers,

You can now monetize your MCP in a few lines of code with @stripe, available today. 💸

- Bill subscriptions or usage-based
- Works with any client
- Supports @Cloudflare's Agents SDK (more to come)

MCP, a new AI-native customer channel. Get started. ⤵️

[Deedy @deedydas]

🚨Viral rumors of DeepSeek R2 leaked!

—1.2T param, 78B active, hybrid MoE
—97.3% cheaper than GPT 4o ($0.07/M in, $0.27/M out)
—5.2PB training data. 89.7% on C-Eval2.0
—Better vision. 92.4% on COCO
—82% utilization in Huawei Ascend 910B

Big shift away from US supply chain.

[Alex Xu @alexxubyte]

Top 6 Tools to Turn Code into Beautiful Diagrams

[Aravind Srinivas @AravSrinivas]

What did we get done this week at Perplexity?

1. Fact-Check any part of the answer with sources: Pick any part of the answer you want fine-grained sources for, or think there's a potential hallucination, and fact-check further.

[Lovis Odin @OdinLovis]

MCP Claude that have full control on ChatGPT 4o to generate full storyboard in Ghibli style ! All automatic I am doing nothing at all, we live a pretty crazy time
@AnthropicAI @OpenAI

[xAI @xai]

This is it: The world’s smartest AI, Grok 3, now available for free (until our servers melt).

Try Grok 3 now: https://t.co/Tj0afLoxEz

X Premium+ and SuperGrok users will have increased access to Grok 3, in addition to early access to advanced features like Voice Mode

[David @dzhng]

Introducing deep-research - my own open source implementation of OpenAI's new Deep Research agent. Get the same capability without paying $200.

You can even tweak the behavior of the agent with adjustable breadth and depth.

Run it for 5 min or 5 hours, it'll auto adjust.

[Mckay Wrigley @mckaywrigley]

A great use case for OpenAI Deep Research is a 1-stop daily news report.

Prompt it with:
- General rules
- Personal bio
- Your interests
- Preferred sources

It’ll generate a comprehensive news report 100% customized to you.

This is how I’ll get my news now.

Full prompt below.

[elvis @omarsar0]

🎓OpenAI Deep Research Guide

Just finished our live webinar on Deep Research, including examples, prompting tips, use cases, and what's missing.

I am releasing the full guide I shared with our members (link in the comments).

[Ruud van der Linden @RuudNL]

Sora v2 release is impending:

* 1-minute video outputs
* text-to-video
* text+image-to-video
* text+video-to-video

OpenAI's Chad Nelson showed this at the C21Media Keynote in London. And he said we will see it very very soon, as @sama has foreshadowed.

[AI Will @FinanceYF5]

这个人帮Uber获得了最初的1亿用户。

然后推动了Tinder、Dropbox和Zynga的爆炸式增长。

他刚刚说：“创作者经济是创新的新前沿。”

Andrew Chen 对未来财富创造的5个预测（以及你为什么应该关心）：

[Star@Day1Global Podcast @starzq]

https://t.co/D8cEcPNPDb

[Thomas Wolf @Thom_Wolf]

You likely missed it if you only follow ML Twitter but there's a series of mind-blowing tech reports and open-source models coming from China (DeepSeek, MiniCPM, UltraFeedback...) with so much lesson learned and experiments openly shared together with models, data, etc

This level of candid sharing of knowledge and insights is something we've lost in most recent western tech models releases and reports (with the noticeable exception of a few places like the recent AllenAI OLMO release)

Just take a look for instance at these two fresh examples published in the past few days:
- the new MiniCPM blog (amazing super small model – deep dive in the experiments): https://t.co/mheQjCkiYW
- the new DeepSeek Math paper archieving over 60% on MATH: https://t.co/MyVKdj9Dsw

[宝玉 @dotey]

年底最值得一读的 RAG 论文：
《Retrieval-Augmented Generation for Large Language Models: A Survey | 面向大语言模型的检索增强生成技术：调查 [译]》

摘要：

在这篇调查中，我们关注的是面向大语言模型的检索增强生成技术。这项技术通过结合检索机制，增强了大语言模型在处理复杂查询和生成更准确信息方面的能力。我们从同济大学和复旦大学的相关研究团队出发，综合分析了该领域的最新进展和未来趋势。

校对中难免有疏漏指出，有翻译错误请指出！

https://t.co/uMmnAbPvHx

[little shrimp🐳 @c_littleshrimp]

币安IEO的 $Arkm, 类似nansen的链上数据分析工具，看起来数据清洗和追踪都更到位，白名单邀请机制。

产品体验链接https://t.co/hWwEOnlm2l

代币经济模型，算是高度控盘了。简简单单，没有套路，上交易所后的第一年因为团队和VC的币还没解锁，流通筹码不多，嗯，适合炒作。

工具里第一个发币的吧

[AI Will @FinanceYF5]

ChatGPT 插件改变了一切。

它的容量现在是无限的。

现在，您可以通过以下方式使用 ChatGPT 与任何 PDF 聊天：

[AlexZ 🦀 @blackanger]

OpenAI 和 吴恩达的提示工程师课程挺给力，学到了一些 prompt 原则和具体策略。做了个脑图。