@cHHillee
@thinkymachines Formerly @PyTorch "My learning style is Horace twitter threads" - @typedfemale
AI
DeepSeek says the model is live on its website, app and API, and optimized for DeepSeekHarness and Tencent CodeBuddy.
AI
In its SFT tests, Tinker says Dimuon often reached lower loss in fewer steps than AdamW.
AI
A user says cached input tokens appear “2x OOM higher than everything else,” followed by two flat lines at zero.
AI
A user names DSV4.1 Flash and GLM 5.3 among the models that currently interest them.
AI
A commenter says scarce data has “absurd and increasing value” as AI models reach narrow superintelligence.
AI
A user who tried Echo before its release says the model did very well imitating both authors' styles.
AI
A user says surprisingly few people have seriously considered why PD disagg is a good idea, and fewer still have compared it with mixed chunk prefill.
AI
Artificial Analysis says GPT-6.1 Sol replaced GPT-6 Sol after seven days and gained four points on its Intelligence Index. Its input and output token prices remain $2/$10 per million, but the cache-read discount rises from 90% to 95%.
AI
A user would welcome a Tinker API for pretraining AI models, arguing that there is no good open-source pretraining stack, unlike for reinforcement learning.
AI
A post seeks suggestions for models missing from Tinker, highlighting surprising omissions and options that would help with specific use cases.
AI
One participant says they see value in managing all the machine-learning infrastructure for users—and that their team is working toward it.
AI
AI benchmarking company accelerates v5 elements via interim index updates with new agent tasks.
Technology
Engineers discuss how dashboards prompt users to count cached tokens.
AI
Researchers note that visible dashboard numbers lead users to include cached inputs when reporting token usage.
AI
Compact multimodal model matches original performance at quarter the size.
AI
Horace He proposes that effective ML techniques increase operational demands on infrastructure teams.
AI
Co-founder announces exit citing health toll from startup demands and highlights key model milestone.
