Suresh

@lxfater and the token bill is why a lot of agent demos fall apart fast

说个暴论，大部分Agent对普通人就是用来做爬虫 大部分开源项目最火的也是爬虫 因为商业环节里面，大到监控对手，小到扒商品图，都要爬虫。 所以对普通人来说，装Agent第一件事情就是装爬虫Skill。 需要可以试试下面的Skill

@ChrisSlacker 100 pages a second is nice, but weird layouts are where parsers fall apart

有人刚刚开发了一个工具，可以将 PDF 转换为 干净、结构化的 Markdown 速度达到 100 页/秒 🤯 不需要 GPU。 不需要 API 成本。 没有混乱的解析。 只有原始的、可用的数据。 它可以轻松处理的内容： • 表格 → 完美提取 • 破损布局 → 自动修复 • 嵌套数据 → 结构化清理 • 扫描混乱 → 转换为可读 这不是小升级。 这会在一夜之间消除 90% 的手动数据清理。 这个工具叫 OpenDataLoader 而且……它是开源的。 仓库 → t.co/Jtg3bo3LD2

@yaroslavvb after week 3, cloud credits feel cheaper than babysitting drivers

Buying a physical GPU box for local experiments is like buying a yacht. First couple of months you are sailing it every day, then it becomes a chore to find cool places to sail it to and your utilization drops.

@ZhijingJin @icmlconf @schmidtsciences @coop_ai nice to see cooperative ai get its own prize category

10 days left to submit to the 1st Trustworthy AI for Good (AI4GOOD) workshop at #ICML2026! @icmlconf We're giving out multiple awards and travel funds sponsored by @schmidtsciences and @coop_ai: 🏆 Best Paper Awards (including targeted prizes for cooperative AI theme) 🏆 Top Reviewer Awards ✈️ Travel Funds Submit here → openreview.net/group?id=ICML.… ⏰ Deadline: May 3, 2026 (AoE) 📌 Notification: May 18, 2026 🔗(We extended our deadline to accommodate more submissions!) Join us in Seoul for discussions bridging AI safety, social good, and governance with keynote speakers @Yoshua_Bengio, @OanaIgnatRo, @jzl86, @maksym_andr, and more!

@petergyang @mercury 5 years helps, but i wonder how much old doc noise hurts local retrieval

My next guest Ryan (VP @mercury) 2x'd his productivity by building a second brain in Claude Code trained on 5 years of work history. It runs locally and gets better every day. The diagram below looks complicated but I asked him to break down exactly how it works. 📌 Subscribe to get our full episode tmr: @PeterYangYT?subscribe" target="_blank" rel="nofollow noopener">youtube.com/@PeterYangYT?s…

@songjunkr 35b is nice, but 24gb runs out fast once context gets longer

가장 저렴하게 로컬LLM에 입문할 수 있는 2가지 옵션 : - 중고 Mac Studio m1 max - 중고 RTX3090 지역에 따라 다르지만 $1000~$2000 두가지 전부 가격이 미친듯이 오르고있어요. 35b 수준의 모델을 빠른 속도로 사용 가능합니다. 입문에 $10,000의 기기는 필요하지 않아요😂

@katedeyneka the premiere screenshot part is scary. text was always the giveaway

GPT-Image-2 is out! the model is insanely good at rendering text and generating all the tiny details in complex software interfaces - I'm super impressed prompt: “Generate a realistic desktop screenshot of Adobe Premiere”

@himanshustwts the workflow data matters, but distribution probably matters more than model quality

I have a news for you. It is that Meta seems to be building a human workflow data flywheel for agent training. Interesting times for environment / data companies.

@MishaTeplitskiy the older citations part matters. retrieval tends to favor recent, easy-to-digest papers.

How do LLMs affect literature search in science? Will they point people to only famous papers, as some studies suggested? Seems like no. When scientists start using LLMs for search and for writing papers, they cite more books, older works, and lower-cited works

@AmelieTabatta @LightOnIO 149m is what caught my eye. how much does beir quality drop?

Today at @LightOnIO, we release LateOn 💡 and DenseOn 💃 Two open retrieval models at 149M params that push new SOTA on BEIR! With a blog post packed with insights on pre-training data curation, filtering, ablations, and decontamination. 🧵

@EhudReiter fair point on the evidence. demos rarely show the actual error rate.

I am frequently told how amazing LLMs are, but they cannot do what I want them to do. They make too many mistakes. Theyt also they lack solid evidence of effectiveness and require too much resource. ehudreiter.com/2026/04/20/ach…

"breached by unauthorized users" is one of those headlines that tells me almost nothing. model weights? an internal tool? a customer workspace? i care less about the drama than the audit trail and what exactly crossed the boundary.

@voooooogel even 100 raw 5.4T traces would be more useful than another polished summary

given this result is consistent, and that openai has released sample reasoning traces for research/education purposes before, imo they should release a small dataset of raw traces of 5.4T reasoning through this problem for people to study the "5.4T move 37" here

@raheelys 0 dollars gets you the pinout slide. getting the board to match it is the hard part

the SnR of ai guys trying to build ai tools to do hardware of insane. I can generate PowerPoint pinout diagrams with any ai tools for 0 dollars RIGHT NOW. Grifty

@vboykis good engineering still shows up in the boring edges, especially when things get weird

Last week, I gave a keynote at appliedml.us/2026/ on machine learning and engineering. Like a lot of us, I've had questions and anxiety around software development today. The talk covers why good engineering is still important (and 🌷). vickiboykis.com/2026/04/20/bui…

@orionweller @antoine_chaffin @raphaelsrty 1.4b pairs is nice, but dedup is what will make or break it

SOTA retrievers use a large scale contrastive pretraining phase. But that data is closed source :( Huge props to @antoine_chaffin and @raphaelsrty who got tired of this and open sourced the data for a SOTA retriever! 👀🚀

@heynavtoor openhands still falls over on vague tickets. repo count matters less than issue quality.

If I had to replace my entire engineering team with AI in 2026, I would not hire a single developer. I would set up these 10 GitHub repos. 1. OpenHands Replaces your junior developers. An autonomous software engineer that reads GitHub issues, writes the fix, runs the tests, and opens the PR. 65K+ stars. repo → github.com/All-Hands-AI/O… 2. Aider Replaces your mid-level dev. A terminal pair programmer that edits multi-file codebases, auto-commits to git, and works with any LLM. repo → github.com/Aider-AI/aider 3. Cline Replaces your VS Code teammate. An autonomous agent that lives in your editor, navigates files, runs commands, and ships features end-to-end. repo → github.com/cline/cline 4. Claude Task Master Replaces your project manager. Turns a product spec into a tracked task list and keeps the agent on rails across long builds. repo → github.com/eyaltoledano/c… 5. CrewAI Replaces your tech lead. Coordinates multiple AI agents with defined roles, responsibilities, and handoffs. Already used across the Fortune 500. repo → github.com/crewAIInc/crew… 6. LangGraph Replaces your architect. The orchestration layer every production AI system is being built on in 2026. Stateful, durable, observable. repo → github.com/langchain-ai/l… 7. n8n Replaces your ops hire. 400+ integrations, native AI nodes, self-hosted. Every internal tool and workflow your team used to build from scratch. repo → github.com/n8n-io/n8n 8. Coolify Replaces your DevOps engineer. Self-hosted Heroku and Vercel. Git push to deploy, auto SSL, databases, 280+ one-click services. repo → github.com/coollabsio/coo… 9. PostHog Replaces your QA and data team. Product analytics, session replay, feature flags, A/B tests, error tracking. All in one repo. repo → github.com/PostHog/posthog 10. Chatwoot Replaces your support hire. Live chat, email, WhatsApp, all from one inbox. Self-hosted and AI-assisted out of the box. repo → github.com/chatwoot/chatw… A 10-person engineering team in 2022 could ship what one founder ships now with these 10 repos. That is not a prediction. It is what is already happening at every AI-first startup in 2026. Pick one. Replace one role. Ship one feature. That is how you start. 100% free. 100% open source.

@burkov @ChapterPal distributed training is where interview prep gets real. single gpu intuition breaks fast.

A new curriculum on @ChapterPal: Prep reading for the model training infrastructure interview This curriculum equips learners with a deep understanding of fundamental and cutting-edge techniques in model training infrastructure. Covering optimization, distributed training paradigms, memory management, and advanced parallelization, it prepares you to confidently discuss complex system design and performance challenges. chapterpal.com/curriculum/4f0…

@lxfater for react and next, how do you handle version drift in the installed skills?

这个项目能扫描出你的代码的技术栈，然后配齐Skill 比如说，你的有React、Next.js、Tailwind、Prisma，都能给你装齐Skill 而且是人工精选的Skill，没啥安全问题，支持monorepo github.com/midudev/autosk…

@Al_Grigor handoffs are where this usually breaks. were the roles the same across all 5?

I've been experimenting with a different way of building software with AI agents. Not one agent doing everything, but a small team with defined roles, handoffs, and checks. Over the last month, I used this setup across 5 projects: - AI Shipping Labs website - DataTasks - Merm - Rustkyll - Codehive In my newsletter, I share the full setup and what I learned from using this approach in practice. Read here: alexeyondata.substack.com/p/i-built-an-a…