Loading activity…
@roberttlange
Staff Research Scientist & Founding Member @SakanaAILabs 🎏 The AI Scientist, ShinkaEvolve, gymnax, evosax Prev: SR & Intern @GoogleDeepMind
AI
Private datasets now carry 45% of the index’s weight, up from 25% in the first version.
AI
The document covers training scale, long-context support, inference optimization and in-house safety benchmarks.
AI
It reportedly tops U.S. open-weight models on TerminalBench 2.1 and runs locally on a MacBook Pro or DGX Spark.
AI
Mistral's report drew praise for technical detail, while another reader questioned model surgery instead of starting over.
AI
A Noise project contributor documents a 22.89-second seed-0 run on one eight-H100 node, down from 123.2 seconds.
AI
Liquid AI pitches a tri-hybrid design, single-GPU fine-tuning and on-device use without detailing hardware requirements.
AI
A contributor reports 3.2115 validation loss, down from 3.4095, after one epoch on a single eight-H100 node.
AI
Mirendil pitches its first model for auditing AI training data and debugging training runs from logs.
AI
A post outlining the work reports roughly 35% lower per-token decode latency with a custom speculative-decoding draft model.
AI
The OPAL team reports gains in selected language and image runs, but most experiments were conducted once.
X
Technology
Mirendil says agents wrote over 99% of the code while a researcher directed the work.
AI
Its developer says Flint exceeds 125 tokens per second on an M4 Pro MacBook Pro at short contexts.
XAI
In the announcement, @bneyshabur places Mirendil-Base in the 100B–500B-parameter range and promises independent evaluations.
AI
Sakana AI says a human-steered early version aided an Error Diffusion study accepted for an ALIFE 2026 oral presentation.
AI
A user says ideas from Nick Cammarata’s interpretability essay inspired the work, which is still in progress.
AI
Reflection says the release includes Mirror CLI, its open-source training harness, and collaborations with 40-plus partners.
AI
Sakana AI says coding agents test hypotheses against preset standards, share results and check which parts of successful programs matter.
