Raktim Mondol

building with ai

[Kexin Huang @KexinHuang5]

Today, we're excited to share that Biomni is published in @ScienceMagazine.

Biomedical research is still fragmented, manual, and difficult to scale. In this work, we introduce Biomni - the first general-purpose biomedical AI agent with an integrated biology environment that can reason, plan, and execute end-to-end scientific workflows.

We show that, with the right environment and harness, AI can automate large-scale omics analyses, orchestrate laboratory robotics, optimize molecular properties, and even train new AI models for biology.

We also introduce a reinforcement learning recipe for continually improving biomedical AI agents, enabling open-source models to achieve frontier-level performance.

It's surreal to look back. We started the Biomni project in early 2024, when agentic AI was still nascent. It is exciting to see tens of thousands of biologists collaborating with agents every day to accelerate science.

Try Biomni: https://t.co/Pzm8TVlkNA
Read more: https://t.co/pRstxzPKDi

This work is not possible without this truly inter-disciplinary team: @serena2z @hcwww_ @YuanhaoQ Minta Lu, Ryan Li, @yusufroohani Lin Qiu @shiyi_c98 Gavin Junze Di @rickwierenga @kavi_deniz Sherry @TianweiShe Shruti Jennefer Xin Zhou @MWheelerMD Jon Bernstein @MengdiWang10 @PengHeAtlas @zhou_jingtian @SnyderShot @lecong Aviv Regev @jure

@StanfordAILab @genentech @phylo_bio @arcinstitute @UW @berkeley_ai @RetroBio_ @tamarindbio @Princeton @UCSF

[关木 @ZeroZ_JQ]

现在最好的agent 框架是什么？

- pi
- open code
- eino

[UI/UX Savior @UiSavior]

Go-to sites for design inspiration

Web design → https://t.co/vy0wk4V0be
Landing pages → https://t.co/DxjvdQ7FG6
Design systems → https://t.co/mjCv4FbpsX
Animation → https://t.co/LZaaK4dbEA
Apps → https://t.co/NYMEfIFcuG
Brands → https://t.co/mUX72HHuxB

[Alex Booker @bookercodes]

Clearest explanations of loops I've read so far

[Eric Topol @EricTopol]

The aging of our immune system is measurable and modifiable. And a youthful immune system is highly correlated with extended healthspan
https://t.co/azMsfd0qfG

[precis0x @precisox]

Te comparto 8 de los mejores repos de github para extraer datos de cualquier sitio web automáticamente:

1. Firecrawl
Pasa una URL y recorre automáticamente todo el sitio, convirtiéndolo en datos limpios y listos para IA. Soporta páginas con JavaScript. Más de 140.000 estrellas.
https://t.co/QzjMTH0M98

2. Crawl4AI
Convierte cualquier web en texto que los LLM pueden leer directamente. Sin API key ni pagos. Lo creó un desarrollador que se hartó de los servicios que cobraban 16 $. Más de 70.000 estrellas.
https://t.co/fzDAMr29W4

3. browser-use
Agente de IA que usa el navegador como un humano: clics, login, formularios, todo. Creado por un startup de ETH Zurich. Más de 100.000 estrellas.
https://t.co/MIvcLE7c8E

4. Crawlee
Base completa para scraping con rotación de proxies, reintentos, anti-detección y manejo de colas. Evita bloqueos de forma profesional.
https://t.co/dQLG9NJdjE

5. Scrapy
El clásico framework de Python para scraping a gran escala. Lleva más de 10 años usándose y procesa millones de páginas de forma estable. Totalmente gratis.
https://t.co/aJYE4Uo1Mk

6. Scrapling
Se adapta automáticamente a los cambios de diseño de las webs y evita detecciones de bloqueo. Funcionalidades de nivel pago… pero gratis.
https://t.co/XyqvZR4vXX

7. AutoScraper
Solo le das un ejemplo de los datos que quieres y aprende solo el patrón. Extrae el resto sin escribir selectores. Muy sencillo y potente.
https://t.co/c6D5FS1ZcX

8. curl-impersonate
Imita las peticiones HTTP de un navegador Chrome real para saltarte las protecciones anti-bot más avanzadas.
https://t.co/dcqagz1uK4

Guárdalos todos, super recomendado.

Con estos puedes sacar datos limpios de casi cualquier web sin pagar servicios caros ni depender de apis limitadas.

[Akshay 🚀 @akshay_pachaar]

You're in an ML Engineer interview at Anthropic.

The interviewer asks:

"Our model generates 100 tokens in 42 seconds. How do you make it 5x faster?"

You: "I'll optimize the model architecture and use a better GPU."

Interview over.

Here's what you missed:

The real bottleneck isn't compute. It's redundant computation.

Without KV caching, your model recalculates the same attention keys and values for every single token generation.

That's why a 9-second inference becomes 42 seconds. You're wasting 80% of your time on repeated calculations.

The fundamental issue:

(refer image below as you read ahead)

LLM token generation is autoregressive:

- Generate token 1 from the prompt
- Generate token 2 from prompt + token 1
- Generate token 3 from prompt + token 1 + token 2

At each step, you're reprocessing ALL previous tokens through attention.

Token 50? You've computed attention for token 1 fifty times.

The reality of attention mechanism:

For each token, the transformer computes:

- Query (Q) from current token
- Key (K) from all previous tokens
- Value (V) from all previous tokens

Then: Attention(Q, K, V) = softmax(QK^T)V

Problem: K and V for previous tokens never change. You're recalculating identical matrices every single step.

How KV caching solves this:

Instead of recomputing K and V matrices:

- Cache them after first computation
- Reuse cached values for subsequent tokens
- Only compute K and V for the new token

Without KV caching (token 50):
- Compute Q, K, V for all 50 tokens → O(n²)

With KV caching (token 50):
- Load cached K, V for tokens 1-49
- Compute Q, K, V only for token 50 → O(n)

You've eliminated quadratic redundancy.

So what's the tradeoff:

While KV caching makes the inference faster, it also takes up a lot of memory, so there is always a tradeoff between speed and memory.

Why your first token always takes longer:

KV caching speeds up inference by computing the prompt's KV cache before generating tokens.

This is exactly why ChatGPT takes longer to generate the first token than the rest.

First token: Computing KV cache for entire prompt

Remaining tokens: Just loading cached KVs + computing new token

----
Everything above is how KV caching works inside a single request.

Running it in production is a different problem. Caches break on document reordering, and a single GPU throws away roughly 15 TB of reusable cache per day.

I wrote an article that picks up exactly where this post ends: how a new open-source architecture manages KV cache at production scale, with 14x faster time-to-first-token to show for it.

The article is quoted below.

[Google Research @GoogleResearch]

Experience AI Scientist, a novel multi-agent system designed to automate the end-to-end scientific research pipeline. Meet Jinsung Yoon and Rui Meng today at 6:00pm at the Google booth (#B206) to learn how complex workflows reliably accelerate AI discovery. #ICML2026

[Yum⋆₊˚ @yuhasbeentaken]

deepseek now has a claude code-style terminal agent...

deep code is an open-source coding assistant built for deepseek-v4.

it supports:

• deepseek-v4-pro and flash
• thinking mode
• reasoning effort control
• agent skills
• mcp integrations
• vscode extension

deepseek moving into the coding-agent layer.

[Nafis Ahmed @iam_nafisahmed]

Asked Claude to design my logo.
The result looked like it cost $500,000.
It cost $0.

The 7 prompts I used (steal these):
Save for later🔖

[Eric Topol @EricTopol]

Using AI to improve cancer immunotherapy outcomes, via training from transcriptomes of 10,000 tumor samples, 33 cancer types @NatureMedicine
https://t.co/Q6UfVQ6wMV

[Scholarship for PhD @ScholarshipfPhd]

Guidelines for writing a research paper for publication (1/5)

[K-Dense @k_dense_ai]

After the Claude Science launch, a lot of people are asking what comes next for agentic scientific workflows. At @k_dense_ai we’re already building there, and we’ll teach you too!
Join us for a free, beginner-friendly workshop on building skills.
Hosted by our AI Engineer @yhhonx , this informal webinar will walk through how to:
☆ Install a ready-made skill from our open-source BYOK skills kit on GitHub
☆ Build a live skill from scratch: a meeting-notes formatter that turns messy notes into a clean, structured summary
☆ Run a real before/after test to see how the skill changes model outputs
To build along, you’ll need:
→ Claude Code
→ Claude Pro or Max
→ An IDE with an integrated terminal, like VS Code or Cursor
→ Node.js 18+
We’ll share the workshop pack in advance, including sample inputs, templates, and prepared before/after reports, so you’re ready to go on day one.
Register here: https://t.co/OXQQZoRO2m

[Moysei @0xMoysei]

Karpathy just wrote the manual for Claude + Obsidian as a real second brain.

Most vaults die the same way. A year of saved articles and highlights. None of it linked. The graph rots while it still looks impressive.

So he moved the upkeep to the model. You curate sources and ask questions. Claude files, links, and reconciles. You keep judgment. It keeps the books.

raw belongs to you and never gets edited. wiki belongs to Claude. It isn't RAG. Your sources compile once into linked pages and compound from there.

9 rules. Start with 10 sources, not 10,000.
Most people hoard notes. This turns them into a brain that maintains itself.

[Raktim Mondol @raktimmondol]

@Voxyz_ai Any difference with graphify?

[Federico De Ponte @fedeponte1]

Introducing OpenPaper: Claude Code for research papers.

I typed one question. Five minutes later, it handed me a complete 8,330-word research paper with 30 citations.

Every citation is a real published paper with a working DOI, verified against 500M+ academic sources. Zero hallucinated references.

18 specialized agents do the research in parallel. Open source.

https://t.co/bqfWZDY9Z5

[Xudong Han @Xudong07452910]

最近看 AI 做研究的论文，感觉一个问题很关键：

Agent 跑得久，不代表它真的在积累研究经验。

这篇 Arbor 论文讨论的就是这个问题。它的核心思路是：让 Agent 不再把每次实验当成一次孤立尝试，而是把所有假设、代码改动、实验结果、失败经验和可复用 insight，都挂到一棵长期存在的 Hypothesis Tree 上。

简单说，就是给 AI 研究过程建一棵「假设树」。

一个长期 coordinator 负责看全局，决定哪些方向要继续扩展、哪些要剪掉、哪些经验可以往上传；短期 executor 则负责在独立 worktree 里实现和测试某个具体假设。

这样每一次实验结束后，系统不是只看分数涨没涨，而是会把证据写回树里，让后面的尝试少走弯路。

我觉得这篇论文最有意思的地方在于，它把 AI 做研究的重点从「跑更多实验」推进到了「让实验经验可积累」。

因为真正的研究不单是不断试错这么简单，而是每一次失败都应该改变下一次怎么试。

对于强的 research agent，可能不只是会写代码、跑实验、刷指标，更要能维护一套长期的研究状态：知道自己试过什么，为什么失败，哪些经验还能迁移到下一个方向。
https://t.co/Sr4lD4pOXT

[Raktim Mondol @raktimmondol]

@gitlawb Your openclude do autocompacting the context window while working

[Raktim Mondol @raktimmondol]

@DavidOndrej1 It is good at Academic Writing. Better than opus and sonnet.

[Amto @XAMTO_AI]

科研圈的天，要塌了！

以前搞研究是啥光景？通宵刷arXiv、反复debug跑代码、一周憋一篇综述，头发掉一地不说，进度还慢得抠脚。

现在Feynman直接一句话给你全包圆：

1️⃣ arXiv深度调研一键到位
2️⃣ 文献综述自动给你吐出来
3️⃣ 代码复现顺手还能验
四个智能体协同干活，幻觉几乎为零，本地跑Ollama数据都不带出电脑的。
说白了，重复劳动全甩给AI，你只管搞创造。研究员、开发者、学生党，谁不用谁血亏。

🔗 https://t.co/rJFGf0xGI3