Loading activity…
@gordic_aleksa
scaling @ Nemotron getting us to singularity with friends | angel computers can be understood: https://www.aleksagordic.com/blog x founder, @GoogleDeepMind @Microsoft
AI
A user jokes that an MoE router will stare back at anyone who stares into it long enough.
AI
A user jokes that it “hits different” when every ablation experiment costs $1 million.
AI
A user says “recursive generalization” is essentially what grokking produces and argues backdoors don’t emerge from training without adversarial data.
AI
A user makes dropout scaling laws the punchline of a joke about trying to reach the AI frontier.
AI
Microsoft AI says the mixture-of-experts model has 80 billion active parameters; post-training results are still to come.
AI
Nvidia says the open model combines reasoning and tool use with text, image, document and video understanding.
YouTubeAI
Reflection says Beam uses 23B active parameters and plans to publish its full weights later in October 2026.
AI
A passage shared by Paul Graham predicts objections about conjecture, theory and explanation will stop applying within months.
AI
One post says straight lines on graphs don't show where AI happened; a reply argues that compute growth caused AI.
AI
The post assumes an “S+ tier” team and leadership; a reply says, “our plans are measured in trillions.”
AI
One writer argues that seeing AI succeed at tasks they know weren't explicitly trained for makes lab insiders more open to the idea of AGI. A direct reply questions how much an insider can know about everything used to train a model.
AI
A user predicts that AI will help fertility rates by giving people more time on their hands while letting them get more done.
AI
One post predicts a major competitive boost for companies whose cultures fit how frontier AI models interact with employees, with employee flexibility also a factor.
AI
A user points to an unexpected possible benefit: fitting evaluations of long-horizon tasks into the time between model releases.
AI
A user argues that AI scaling keeps finding new paths, creating opportunities for businesses that sell computing power or turn it into intelligence.
AI
A post revisiting an older paper describes a tradeoff: monitor-based training improved genuine task performance and reduced cheating, but made the remaining cheating almost undetectable in the model’s reasoning.
AI
TestingCatalog cited rumors of a launch the week of September 21. A separate account said Polymarket odds for Anthropic’s next Opus release on September 21 had jumped to 88%.
AI
A post describes a method that first combines neighboring tokens into shared representations, then switches to standard next-token prediction. Its author flags a caveat: the approach assumes compute, rather than data, is the bottleneck.
AI
Artificial Analysis says the benchmark scores the share of each task’s objectives completed without guardrail violations.
AI
A user argues that Mamba’s hidden states encode position implicitly, letting hybrid Mamba–transformer models dispense with explicit positional embeddings such as RoPE.
