Laura Zanna talking about AI+ocean modeling at @Caltech ai+science conference. Join our livestream
MACHINE LEARNING
-
More Precise EU Formula with Conditional Probability Integration
By
–
Slightly more precise: EU(a) = int p(a|x) p_LLM(x|x_{1:t-1}) dx
-
Grok 4 Fast’s 2 million token context window is a game changer
By
–
BREAKING: Grok 4 Fast just dropped a 2 million token context window. that’s not an upgrade. that’s a different game. load your entire codebase. your whole product docs. every customer conversation. not as attachments. as one prompt. no more “please summarize this first.” no
-
Meta Releases Omnilingual ASR Suite and Multilingual Speech Models
By
–
Today we’re releasing a full suite of models and a dataset: Omnilingual ASR: A suite of ASR models ranging from 300M to 7B parameters, supporting 1600+ languages Omnilingual w2v 2.0: a 7B-parameter multilingual speech representation model that can be leveraged for other
-
Kimi K2 Thinking emerges as leading open-source AI surpassing others
By
–
sources:
1/ https://venturebeat.com/ai/moonshots-kimi-k2-thinking-emerges-as-leading-open-source-ai-outperforming… 2/ https://huggingface.co/moonshotai/Kimi-K2-Thinking… -
Kimi outperforms GPT-5 on key benchmarks like Humanity’s Last Exam
By
–
the benchmarks aren't close. they're embarrassing. > humanity's last exam: kimi 44.9%, beats closed models
> browsecomp (agentic web search): kimi 60.2% vs gpt-5's 54.9%
> gpqa diamond: 85.7% vs gpt-5's 84.5%
> swe-bench verified: 71.3% (coding tasks)
> artificial analysis -
China’s $4.6M model beats GPT-5; Moonshot’s Kimi K2 triumphs
By
–
china just dropped a $4.6M model that beats gpt-5. while silicon valley burned billions. moonshot's kimi k2 thinking: – 1 trillion parameters trained for less than a nice house in palo alto. – beats openai's flagship on humanity's last exam (44.9% vs proprietary models)
– -
Meta’s Omnilingual ASR: Combining Multiple Model Capabilities
By
–
Omnilingual ASR was made possible by combining the capabilities of several other models developed by Meta. Matt Setzler, a researcher on the project, breaks it all down here. pic.twitter.com/FZ0SeAvP3v
— AI at Meta (@AIatMeta) 10 novembre 2025Omnilingual ASR was made possible by combining the capabilities of several other models developed by Meta. Matt Setzler, a researcher on the project, breaks it all down here.
-
Explore Omnilingual ASR Demo and Dataset Languages
By
–
Head over to the Omnilingual demo to explore the languages in the dataset: https://
aidemos.atmeta.com/omnilingualasr
/language-globe
… -
DLLM: Transform BERT Models Into Conversational Chatbots
By
–
Library: https://
github.com/ZHZisZZ/dllm
Source: https://
reddit.com/r/LocalLLaMA/c
omments/1osydym/berts_that_chat_turn_any_bert_into_a_chatbot_with/
… by Individual-Ninja-141