Good catch! This likely would be fixed with more robust training (more steps + rollouts, and ideally a larger than 1.5B model) 🙂 This was mostly just to show how it works. Feel free to try the same task with a larger model, more steps etc.
RESEARCH
-
Expanding Context Length to Human Lifetime Audio Scale
By
–
"there is no reason we can't make the context length 100 words, which is roughly what a human hears in his lifetime. There is no reason we can't do that. Its really inference support."
-

Dario Amodei on Exponential AI Growth and Scaling Laws
By
–
New interview with Dario Amodei, here is the tl;dr – Exponential growth continues: Amodei firmly believes that the exponential increase in AI capabilities will continue.
– No end in sight: He currently sees no end or slowdown to this trend. The “scaling laws” (more computing -
RL Model Training Made Accessible for Developers
By
–
You literally can’t be too lazy to train a model with RL with this approach No excuses anymore Go build!
-
China’s AI Development Relies on Student Labor and Public Data
By
–
https://
restofworld.org/2023/china-ai-
student-labor/
… and https://
scmp.com/economy/china-
economy/article/3294867/chinas-ai-aspirations-boosted-public-data-allowed-labelling
… are two relevant stories -

Promoting Responsible AI Engineering Against Hype and Overpromising
By
–
another important agenda i had was to push "responsible" ai engineering forward in as mass appealing a way as possible – a very difficult thing to do in a world of extreme hype tweets and flashy overpromising overproduced videos. https://
x.com/aiDotEngineer/
status/1950635169133797517
… so honored/happy to have -

Multi-Agent Systems: Honesty in Prompting and Post-Training Efficiency
By
–
(i'm just memeing, no harm meant, multi agents is obviously the way, per https://t.co/dYMgFl7bFw , however 1) you should be @lateinteraction -level intellectually honest about your prompting and performance gains and 2) multiagent posttraining is going to be ooms more efficient…
— swyx 🐣 (@swyx) 30 juillet 2025(i'm just memeing, no harm meant, multi agents is obviously the way, per https://
x.com/latentspacepod
/status/1939848051847491652
… , however 1) you should be @lateinteraction -level intellectually honest about your prompting and performance gains and 2) multiagent posttraining is going to be ooms more efficient -
Researchers solve hardest problems so engineers focus on implementation
By
–
Researchers solve the hardest problems so engineers don't have to.
-
Scientific AI Integration in Research Databases and Workflows
By
–
5. The tool is deeply integrated into scientific databases and workflows – from PubMed to arXiv. This makes the answers accurate, citable, and directly reusable.
— Chubby♨️ (@kimmonismus) 30 juillet 2025
Scientific AI should not be a gadget, but a tool that saves time and speeds up thought processes. That's exactly what… pic.twitter.com/0KNcFetHgH5. The tool is deeply integrated into scientific databases and workflows – from PubMed to arXiv. This makes the answers accurate, citable, and directly reusable. Scientific AI should not be a gadget, but a tool that saves time and speeds up thought processes. That's exactly what
-
SciSpace Agent: Automating Research Review and Data Analysis
By
–
4. Example: You are working on a review of neuroplasticity. SciSpace Agent scans current studies, generates an outline, provides precise citations with DOIs, and even writes suitable abstract drafts—all in seconds.
— Chubby♨️ (@kimmonismus) 30 juillet 2025
Or you are analyzing complex data: SciSpace can help you… pic.twitter.com/qJl6IKLPaJ4. Example: You are working on a review of neuroplasticity. SciSpace Agent scans current studies, generates an outline, provides precise citations with DOIs, and even writes suitable abstract drafts—all in seconds. Or you are analyzing complex data: SciSpace can help you