Loading activity…
@arankomatsuzaki
Sharing AI research. Early work on AI (GPT-J, scaling, MoE). Ex ML PhD (GT) & Google.
AI
Nace AI’s upgraded decision model runs locally on Macs via Kapitsa, according to arankomatsuzaki.
AI
Drex reportedly assigns probabilities to fixed-choice answers instead of generating text.
AI
LiberAI's model reportedly completed 36% of 600 attempts across 60 new tasks without demonstrations of those tasks.
AI
The reported MATH-500 score rose from 7.9% to 48.1% with the same sampling settings.
AI
OpenBMB describes Chinese and English document support; Jev scores candidate OCR results for text, formula and table regions.
AI
OpenAI says its claimed refutation of the Exponential Time Hypothesis was formalized in Lean and made public.
AI
A user exploring the idea asks whether anyone is interested in cofounding or knows someone who might be.
AI
One commenter says AI mostly helped with rewording; their feed split between “unreadable slop” and praise.
AI
A post describes a benchmark testing nine frontier agents on 15 open scientific follow-up problems developed with domain scientists.
XAI
Quoting a post captioned 'recession indicator,' a user said it explained what was exciting about GLM 5.3 Flash fine-tune.
AI
A post links to a photo without offering any explanation beyond 'hmm.'
AI
A post claims a more than 100-fold question-answering speedup for VideoJev versus Thinking.
XAI
A post describes stage-by-stage checks that send a failed tablecloth grasp back to repositioning.
AI
A post claims ARC achieves more than 5× higher success on reasoning-heavy tasks without changing the model architecture.
AI
The ARC project says it adds action-grounded reasoning to existing models without new robot demonstrations or foundation-scale training.
AI
A user floats buying a few hundred robots in India for companies to run real-world tests.
AI
Partiful’s listing invites researchers, engineers and innovators to commemorate mathematics and discuss scaling laws.
AI
A post introducing the report claims three years of data support its argument about AI and creativity.
XAI
A post says success on RoboCasa GR-1 rose from 63.3% to 78.7% as context grew from 0.0 to 19.2 seconds.
XAI
A post describes an 8B-parameter RoboJEPA model from Meta’s FAIR and Mila, trained on 15,022 hours of robot video.
