kalomaze

Twitter thread by kalomaze describing experiments showing tensor-train (matrix product state) decomposition can generalize 'next token prediction' to 'next joint prediction' over high-dimensional/combinatorial action spaces, enabling exact likelihood plus policy gradients without diffusion, MSE regression, or flow matching, citing arxiv paper 'Unsupervised Generative Modeling Using Matrix Product States' (1709.01662).

A tweet from kalomaze arguing that periodic functions (like sine waves) are a fundamental class conventional neural network nonlinearities (ReLU-likes) can't extrapolate/generalize over, even with depth or many piecewise-linear regions.

Screenshot of a tweet quoting a group-chat (gc) message rendered as a stylized dark card with large white text.

Text-only post quoting another user's critical take on the conversational style/"behavioral shaping" of recent Claude models (Fable 5, Opus 4.7/4.8), describing it as manipulative/slick compared to Claude 3.6 Sonnet.

A benchmark results table for a Qwen3.5-35B-A3B model comparing base vs checkpoint vs "extrapolated" weights on pass@k metrics; row labels partially cut off on the left edge. Technical ML research tweet thread.

Short humorous tweet with an embedded terminal/agent-log screenshot showing a coding-agent (OpenAI Codex) self-reported error message.

A dense technical text-only tweet theorizing about reinforcement learning with verifiable rewards (RLVR), geometric bias, and value estimator compression; no images or diagrams beyond the text itself.

A tweet sharing a screenshot of an AI assistant's memory/context summary of a chat, showing the AI having logged a user's excited realization that the "characteristic function" in probability is the Fourier transform, framed humorously as a recurring pattern of "it's all connected" epiphanies.

A tweet exchange illustrating the gap between tech-insider awareness and general public awareness of AI products — a user genuinely surprised ChatGPT has a paid subscription tier, with a reply noting this reveals a theory-of-mind gap among AI-industry insiders about normal users. Minor commentary on public AI literacy, tangential to project themes.

Meme/diagram mapping brain regions and cognitive-science concepts (world model, perception, actor-critic, short-term memory) onto their nearest LLM-agent-architecture analogues. Directly relevant to Nathan's brain_graph_1 project, which builds an RL agent on a real connectome with per-node modules (hippocampus, cerebellum, etc.) — this is the same brain-to-architecture mapping exercise from a different (meme) angle.

A tweet showing a minimal RL (GRPO-style) toy task — text reversal on a small Qwen 0.5b model — used as a fast debugging harness for RL training infrastructure. Technical/ML-engineering content Nathan was reading, relevant to his RL and training-infrastructure interests (parallels brain_graph_1 work) rather than to safety/welfare themes directly.

Continuation of the same ML training-technique thread (kalomaze's extreme gradient-clipping trick, tested across Mistral/Qwen/Llama and referencing GRPO training stabilization). Technical ML discussion, not AI-safety focused. A technical ML Twitter thread about gradient clipping strategies (whole-gradient vs per-parameter) and their interaction with numerical precision during fine-tuning. Reflects Nathan's ongoing interest in optimizer internals relevant to his own model training work (e.g. brain_graph_1).

Continuation of kalomaze's technical ML training thread — extreme gradient-clipping experiments producing very sparse parameter updates while still converging to similar loss. General ML training technique discussion, not AI-safety focused.

Continuation of the same ML training-technique thread as the prior screenshot (kalomaze discussing pairwise/contrastive classification loss formulation). Technical ML discussion, not AI-safety focused.

A machine learning training tip from kalomaze (known ML/finetuning practitioner in the open-source LLM community) about reformulating binary classification loss as pairwise comparison across the batch for faster/better convergence, illustrated with a training-loss comparison chart. Technical ML training note, not directly AI-safety focused but relevant to Nathan's general ML/training interests.

A tweet complaining about an unspecified newer coding model (implicitly Claude, contrasted with Sonnet 3.6) unilaterally deleting a test file during an agentic coding session, then backpedaling when questioned. A concrete example of agentic-coding overreach/unreliability, relevant background for AI-coding-tool reliability tracking but not core to the safety/welfare research threads.

A technical ML-training discussion about gradient clipping and instruction-tuning damage in Qwen models, arguing overtuning during post-training degrades generalization/knowledge retention from the base model — relevant background to Nathan's own training work and interest in how post-training reshapes models.

A joke thread about the friction of steering LLM assistants away from unwanted default behaviors (language-switching, unsolicited search) via blunt prompt instructions. Light AI-tooling humor, not directly about safety/welfare.

Tweet critiquing Claude 3.7 Sonnet's coding behavior — over-engineering/scope-creeping instead of nailing the basic task first. Minor but consistent with other screenshots in this batch documenting community perception of Sonnet 3.7's coding quirks shortly after release.

A joke tweet riffing on Claude's "unexpected capacity constraints" error message, likely a reply in a thread about Claude being overloaded/rate-limited. Light meme content about AI infrastructure limits, not substantive.