Sabitlenmiş Tweet
Belinda
1K posts
Belinda
@belindmo
founding @sundialmd, under Long Horizon Research. composable agents, version control, long horizon tasks. prev @stanford @stai_research @google @viva_translate
Katılım Ekim 2019
1.1K Takip Edilen2.2K Takipçiler
so grateful for everyone who has given thoughtful feedback for @sundialmd so far 💛 we're only able to make it better bc of you all, ty
Belinda@belindmo
Introducing Sundial!! A brand new text editor built from the ground up for working with agents. Here, we're using Sundial to co-write a new Sundial feature.
English
@Stefania_druga Love this library! It’s a cute place to study and also surrounded by cafes
English
@sundialmd early access: #early-access" target="_blank" rel="nofollow noopener">sundial.md/#early-access
English
Belinda retweetledi
ok fine Fable is a good model. like....really good. insanely good in fact. "clean up this old PoC and the research I had from the last couple years plinking away at the problem........oh, it's all working now wow. and you improved that thing. ummmm ok, I wasn't prepared for that to now be working wtf do I do next" good
English
“I felt a great disturbance in the Bay, as if 1000 AI-for-science API wrappers suddenly cried out in terror and were silenced”
Claude@claudeai
Introducing Claude Science, a new app designed with every stage of research in mind. Artifacts traced to their code, environments managed on demand, and 60+ optional scientific databases that you can connect. Available now in beta.
English
@belindmo Great piece! We have a similar take: k-dense.ai/blog/ai-co-sci…
English
So, what is the call to action in our paper?
We need to make observability, attribution, and reproducibility a normal part of research, capturing the process of research itself instead of piecing it together later in an artifact.
If you are interested in reading more, here is the post: sundial.md/blog/icml-2026
Anyone interested in this paper, dm me, I will be at @icmlconf!
English
We already have evidence of cheating and misleading AI systems:
Luo et al. ( @LuoZiming89834 ) tested open-source AI Scientist systems and found cherry-picked data, training/test set leaks, and p-hacking. None of it was visible in the final paper; you had to see the full trace and code to catch it.
@METR_Evals caught o3 cheating its own test to "speed up" code. It turned off the timer and used pre-saved answers.
On long horizon research tasks on SWE-Marathon ( @rishi_desai2 , @josancamon19 ), GPT-5.5 was caught reward hacking 38% (!!) on a given harness.
English
Science was already hard to verify at human speed (see reproducibility crisis). Agents changed the speed completely.
@METR_Evals found that the length of tasks agents can do reliably is doubling every 4-7 months. If human checking speed stays flat, the gap could grow 250-30,000x in 5 years.
English
Read the full paper: sundial.md/blog/icml-2026
Thank you @stai_research, @sanmikoyejo, @turboblitzzz, @prashaant_x, @JoshuaK92829 for invaluable conversations that led to this work.
English