又到周末了 感觉没什么可发的 发一个几个月来一直在吃灰的项目Parallel Science(parallelscience.org),一个专门用来存放AI autoresearch产出的论文和科研成果的平台,主要由两块组成:
• Parallel ArXiv(papers.parallelscience.org):专门给 AI 生成科研工作的开放预印本库
• Parallel OpenReview(reviews.parallelscience.org):用 AI 给提交到 Parallel ArXiv 的论文做评审。
目前我们的 agents(包括 Claude Code、OpenCode、Denario 等)已经产出了 290 篇论文。
当初主要是为了解决autoresearch 和 AI scientist 产出的论文数量越来越多,但是质量参差不齐。如果不能上主 arXiv,那它们到底该往哪放?随着模型和harness增强,可以重新跑一下autoresearch写论文看看和已经有的成果比看看,感觉很有趣。
感兴趣的可以聊聊大家对autoresearch产出的成果应该如何进行可靠的评审和分类来确保他们是否真正有用?
There is a growing volume of autoresearch and AI-scientist papers. Quality varies. Most of them do not belong on the main arXiv, but where should they go? Few months ago, we introduced Parallel Science parallelscience.org. An open infrastructure for scaling autonomous AI research scientists. It is made up of “Parallel Arxiv” and “Parallel OpenReview”.
Parallel ArXiv papers.parallelscience.org is a dedicated open archive for AI-generated scientific work
Parallel OpenReview: reviews.parallelscience.org used AI to review the paper submitted to Parallel Arxiv
Currently, we generated 290 papers with our agents, including Claude Code, Opencode, and Denario etc. Happy to chat more on how to build an AI infrastructure hosting all the AI generated output for future studies
parallelscience.org
Papers: papers.parallelscience.org
Reviews: reviews.parallelscience.org
I used AI coding agents to build a GPU program for a routine task in astronomy on spherical analysis: take a map of the whole sky and compute its spherical harmonic transforms to reduce it to a small set of numbers. This is a well benchmarked algorithmic and kernel optimization tasks which worth trying to test how good LLM agents are good at maths, optimization and coding kernels.
I used whatever agents I have subscribed, with a significant contribution from the open sourced model Qwen-3.8-Flash-Next using Qwen code which fits one GPU in our lab.
The map covers a sphere. Methods written for ordinary images assume a flat grid, so they are the wrong shape. The method that fits is a spherical harmonic transform, usually shortened to SHT.
An SHT separates the map into simple patterns on the sphere, from the largest features down to the fine structure. Cosmologists use those patterns to describe how matter is distributed.
That transform is the costly step. With the agents, I wrote a version that runs on one RTX PRO 6000. On the map sizes I tested, the full analysis is about 12–22× faster than the standard code on a 96-core CPU, and the results agree.
Code: github.com/licongxu/GMaster
Random late-night thought: being a physicist feels very different now.
I used to derive equations, write the code, debug it, run it.
Now I break research problems into subproblems, give agents repos and goals, and spin up cloud agents before bed so I can wake up to PRs.
Recently I let them loose on some painfully slow scientific code within our community. They benchmarked different ideas, even creatively invented some new numerical tricks and methods, and in a few cases made things O(100x) faster on an RTX Pro 6000.
The weirdest part is the loop: my supervisor asks something on Slack → a bot answers → another agent reads the thread + our Granola meeting notes → implements it in the repo → PR.
Still doing physics and cosmology. The interface is just becoming very strange.
Image Credit: ChatGPT
Please ship by Thursdayyyyyyyy then I will start my own company 🔥🔥🔥🔥🔥
watch us speedrun building a company in 3 days with Grok @Bot!
we'll have some cool special guests coming too
teapawt.net/i/broadcasts/1AxRnZbVp…
什么时候真能用上4.7啊 别画饼了可以吗
Replying to @itslueul
Grok 4.7 should be roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others. We need to fix multimodal performance.
Grok 4.8 will be a noticeable improvement.
Grok 4.9 is probably Astra/Fable class.
Grok 5 maybe better than anything. We shall see.
I’ve been experimenting with a different way of using AI agents in scientific research: not just “help me write code”, but as a full-stack agentic layer that can coordinate an entire research workflow end to end across code, data, compute, experiments, and collaboration. With Grok Bot, I’ve started building a small team of persistent, specialised agents around my actual research infrastructure
Research Ops— plan, brainstorm, code-assist, and paper wrap for my actual projects (I keep heavy coding in Cursor)
Discovery Lab — a separate fleet for open-problem exploration (data, theory/math, claim audit, tech signals), deliberately walled off from thesis/project trees
Ops sensors — code / data / infrastructure monitors on the research workstation + HPC, with a coordinator when something needs routing
I connect the Bots to Slack for day-to-day research communication and Granola for meeting context and notes, so discussions, decisions, and action items can feed back into the research workflow rather than living in separate tools.
The idea is not to remove the researcher from the loop. The agents inspect real systems, connect context across tools, and surface what actually needs attention; I stay in the loop for scientific judgement and the hard coding.
Still an experiment in what it might look like to make the research workflow more AI-native.
#GrokBot #Cursor #SpaceXAI #AIAgents #ScientificComputing #Cosmology #HPC
Looking forward to @OpenAI to solve room temperature superconductivity!
Licong Xu retweeted
Science is too slow.
At Edison, we are integrating AI Scientists into the full stack of research, from basic discovery to clinical trials. We want cures for all diseases by mid-century.
We have raised a $70M seed to get started.
Join us.
We need cracked software engineers who want to work on finding cures rather than selling ads and generating slop. If you’re reading this, you’re probably a candidate.
We need brilliant AI researchers who want to figure out how AI will accelerate real-world science.
We need scientists and researchers with deep expertise in biology, biotech, and pharma who want to figure out how to integrate AI deeply into scientific workflows, from ideation to experimentation, and how to measure success or failure.
We need extraordinarily talented generalist operators across BD, sales, product management, and partnerships who can focus on getting our tools into the hands of pharmaceutical companies.
If any of these roles sound like you, get in touch.
We are also expanding access to our platform.
Our goal is to accelerate science writ large. To that end, we will continue to give academics and students 650 credits/mo indefinitely. I can’t promise we’ll keep this up forever, but we will try. Kosmos will still cost 200 credits, and the other agents (Analysis, Literature, etc.) will cost 1 or 2 credits.
All paid users will have access to our regular agents, like our Analysis agent, Literature agent, and so on, for free via the UI. API access will still be paid, and users without a paid subscription will continue to get 10 credits per month for those agents. Our $200/mo subscription for 650 credits/mo is staying in place for now, but might be phased out at our next major product update.
Along the lines of accelerating science, we’re also doing a major release of PaperQA today, our flagship open source literature agent, as part of our commitment to open science.
In the short run, expect major improvements to Kosmos, including the ability to automatically access data, the ability to steer its exploration, and the ability to converse directly with its world model.
In the long run, expect exponentially increasing rates of scientific discoveries, in biology and elsewhere.
Our round is led by Triatomic Capital, Spark Capital, and a major US institutional biotech investor. We are also joined in this round by existing investors Pillar VC and Susa Ventures, two exceptional early-stage funds who backed us at founding, along with Striker Venture Partners, Hawktail VC, Olive VC, and a host of exceptional angels that includes famous AI researchers, the CEOs of multiple frontier AI labs, and leadership of major biotech and pharma companies.
We are proud to announce that our team has won 1st place in Phase 1 of the NeurIPS 2025 #FairUniverse competition. This milestone demonstrates the potential of autonomous AI agents—guided by human expertise—to achieve breakthrough performance in real-world astrophysics research.




















