Pinned Tweet
@Nature published an article on our project, data-to-paper
Scientists used ChatGPT to generate an entire paper from scratch — but is it any good? go.nature.com/3rjCrc1
Tal Ifargan retweeted
With my EPFL pal @GiorgosNik02 and the @GladiaLab (@DonatoCrisosto1 @teelinsan @EmanueleRodola ) we proved that LLMs are injective (in the hidden representations) and invertible!
My fav takeaway: information is preserved along the residual stream throughout the transformer.
Tal Ifargan retweeted
Original Article by @TalIfargan et al.: Autonomous LLM-Driven Research — from Data to Human-Verifiable Research Papers nejm.ai/4f8JAPA
#ArtificialIntelligence
ALT Figure 1. Data-to-Paper Orchestrates Agent Interactions and Information Flow through a Multistep Research Cycle from Annotated Data to Research Paper.
ALT Figure 2. Exchange among LLM Agents Helps to Refine Research Products.
ALT Figure 3. Data-to-Paper Is Able to Autonomously Create Correct Papers for Simple Research Goals and Datasets While Human Copiloting Is Required to Ensure Accuracy in More Complex Settings.
Indeed, mistakes in research are inevitable, but the key is to leverage AI to address them more effectively and expose them transparently. We should harness AI to make science more reproducible and verifiable.
Tal Ifargan retweeted
A study by @TalIfargan and colleagues demonstrates a potential for AI-driven acceleration of scientific discovery in biomedical research and beyond, while enhancing, rather than jeopardizing, traceability, transparency, and verifiability. nejm.ai/4f8JAPA
Tal Ifargan retweeted
AI to drive to research goals from data to publication? A glimpse of which fraction of biomedical research? @NEJM_AI ai.nejm.org/doi/10.1056/AIoa…
Tal Ifargan retweeted
Agree with this and it’s not so different from the folks who complain AI will break something in science or the practice of it (usually something that’s already deeply broken or aggressively suboptimal)
The people freaking out about AI "destroying education" are telling on themselves. They're so indoctrinated by our Industrial Revolution-era factory model of education that they can't imagine alternatives. But every major technology - from books to electricity to the internet - has transformed how humans learn.
If AI can do your homework, maybe homework isn't how we should measure learning anymore. The goal of education is to create lasting cognitive change - literally rewiring neural pathways as we master reading, writing, and analytical thinking. Our current system of standardized assignments and generic grading is just one deeply flawed path to that destination.
When someone says "AI is great for learning but will destroy education," what they're really admitting is that our education system is broken and desperately needs disruption. The fact that so many educators can't envision alternatives proves how effectively this system has educated the creativity and vision right out of us. Time to break free.
Tal Ifargan retweeted
Our data-to-paper AI is officially out!
ai.nejm.org/doi/10.1056/AIoa…
Autonomously analyzes data and writes human-verifiable papers!
Fun AI-human copiloting!
Papers are "data-chained": click-traced to the raw code.
Try it out:
github.com/Technion-Kishony-…
@TalIfargan @LostInTranscrip
Tal Ifargan retweeted
AI-Science: New features!
- Human-AI copilot app: Much more powerful (and fun) than the AI agents doing it all.
- Fully annotated figures!
- “Data Chaining”: All results (inc Figures) are click-traced to the specific code lines that created them!
pip install data-to-paper
Tal Ifargan retweeted
学术福音🔥 这项 「Data to Paper」的研究有点炸裂🤯🤯
一套包含 9+过程 LLM + 规则驱动的 Agentic Workflow 组合，实现从数据探索到提出研究问题、实现研究目标、实验、写论文+引用的端到端全流程 ⚡️
甚至提供一套人机交互的 UI 辅助做交验，提高可靠性！
源代码地址 👉 github.com/Technion-Kishony-…
And we’re constantly improving it! 🚀
Recently added:
- The ability to reset to a designated step
- Monitor API usage costs
We’re also working on a major feature we’ll announce soon! Stay tuned and keep engaging with us: github.com/Technion-Kishony-…
Tal Ifargan retweeted
Revolutionizing scientific research with AI! Introducing #data-to-paper by #Technion-Kishony-lab - a groundbreaking framework that automates the journey from raw data to a comprehensive scientific paper.
github.com/Technion-Kishony-…
Tal Ifargan retweeted
Kind of a déjà vu of "GPT-1 of AI science” moment though...
github.com/Technion-Kishony-…
The AI Scientist discussed in @Nature‼️
“All my colleagues in different sciences are trying to figure out, where does AI fit in in what we do? It does force us to think what is science in the 21st century—what it could be, what it is, what it is not.”
nature.com/articles/d41586-0…
Tal Ifargan retweeted
welcome to the party guys
Fully automated LLM driven science was first showcased more than a year ago ...
nature.com/articles/d41586-0…
github.com/Technion-Kishony-…
arxiv.org/abs/2404.17605
I recommend the comprehensive list posted by @Hoper_Tom :
xitter.cf/Hoper_Tom/status/18232…
Amusing you use a "self-assessed novelty checker", while failing to mention *a ton* of work. Here, "AI Scientist", some data to improve your novelty checker:
xitter.cf/roykishony/status/1785…
xitter.cf/mbodhisattwa/status/18…
xitter.cf/_DougDowney/status/174…
xitter.cf/Hoper_Tom/status/17599…
xitter.cf/jinheonbaek/status/177…
The space is getting crowded
welcome to the party guys
Fully automated LLM driven science was first showcased more than a year ago ...
nature.com/articles/d41586-0…
github.com/Technion-Kishony-…
arxiv.org/abs/2404.17605
I recommend the comprehensive list posted by @Hoper_Tom :
xitter.cf/Hoper_Tom/status/18232…
Can LLMs perform research?
In this article, we share the conclusions we gathered throughout our journey working on the data-to-paper framework. We also introduce a handy "how-to" section to help you kickstart a research project with your dataset using data-to-paper.
our thoughts on the place of AI in Science...
And a getting-started guide to "data-to-paper"
medium.com/@talifargan/data-…
Remember this dataset that you have, that you always said you have to find time to sit down and analyze? Now, LLM agents can assist you! We have built a GUI app for #datatopaper that allows easy data uploading, overseeing and reviewing. Check it out: github.com/Technion-Kishony-…
Tal Ifargan retweeted
1/n Can LLMs perform scientific research? And, can they do so while enhancing key scientific values including transparency, traceability & verifiability? Check out our preprint on the "data-to-paper" platform. w the amazing @TalIfargan @LostInTranscrip arxiv.org/abs/2404.17605
Tal Ifargan retweeted
"שאולי ראית שביבי מחק את הציוץ שלו?"
"בוודאי שראיתי אמציה, שגיאה קשה"
"אין ספק, בימים שכאלה לפרסם ציוץ שכזה?"
"אל תהיה דביל, שגיאה שהוא מחק. הוא התנער מהסרת האחריות, זה חמור מאוד אמציה"
"אה שאולי, אני מצפה ממנהיג אמיתי שיידע לקחת אחריות גם במלחמה"
"ואני מצפה משרון סטון להכין לי טוסט גבינה, אבל היא צמחונית. באמת אמציה, אם היית פותח פה ושם ספר היסטוריה היית יודע שמנהיג אמיתי מחובתו לברוח מאחריות. נראה לך ג'ינגס חאן צייץ שהוא לוקח אחריות על הכישלון מול אבבא? כשוושינטון פתח מלחמה של usa מול אנגליה וכמעט איבד להם את כל התה וכולם זעמו עליו נראה לך הוא התנצל? לא, הוא ברח מאחריות עד הניצחון והיום יש על שמו כיכר בניו יורק"
"אבל שאולי, זה לא המאה ה-18 יותר, אנחנו בעידן אחר"
"אמציה, לקחת אחריות זו חולשה. אתה יודע שבני אדם זו החיה היחידה שלוקחת אחריות בטבע? ראית פעם בנשיונל ג'וגרפיק מנהיג של שימפנזים מתנצל כי הוא בטעות החליק על בננה והרג זברה? ראית כריש מתפטר מתפקידו? אפילו אריות מתחמקים מאחריות וכל היום ישנים. האריה היחיד שלקח אחריות פומבית היה מופאסה ובסוף הוא נהרג בשוגג, לכאורה. רק מנהיג לא אחראי לוקח אחריות"
"אני לא יודע, אולי אתה צודק"
"הדבר האחרון שאנחנו צריכים זה מנהיג נושא באחריות. תראה את סינוואר לאילו הישגים יפים הוא מביא את חמאס במלחמה, ולא אכפת לו בכלל שכל העם שלו ימות מצמא בלי נטפליקס באוהל"
"כן אה"
"אבל אני אופטימי, בעזרת השם המלחמה רק התחילה ויש לביבי עוד הרבה אחריות שהוא יכול לא לקחת, הוא רק צריך להיות מספיק אמיץ"
