The Archive

Google CEO Sundar Pichai arrives for the inauguration of President Donald Trump. | Getty Images Over 600 Google employees signed a letter to CEO Sundar Pichai demanding that Google block the Pentagon from using its AI models for classified purposes, reports the Washington Post. Its organizers claim many of the signers work in Google's DeepMind AI lab, and include more than 20 principals, directors, and vice presidents. According to the Post, the letter says that "The only way to guarantee that Google does not become associated with such harms is to reject any classified workloads. Otherwise, ...

I'd jump on runpod and ssh in to test my workloads, but they don't have it. Would love to know how well this runs, particularly as context approaches a full 256K. Thanks!

The unwinding of Meta’s deal shows how tech founders struggle to cut China ties.

MIMO V2.5 PRO model released on Hugging Face; limited details available from Reddit post.

Pattern-based KC extraction from student code to auto-generate personalized worked examples for programming education.

Closes sample complexity gap for multiclass classification via DS dimension algebraic characterization.

Reddit user reports Claude helped manage ADHD-related decision paralysis and life organization tasks over one day.

Reddit user shares MCP tool that aggregates SEC filings (Form 4, 8-K, 13D) for real-time insider trading data.

HRGrad method resolves gradient conflicts in multi-task learning across asymptotic kinetic regimes.

Learning from multi-source CoT supervision can be cryptographically hard despite easy single-thinker supervision.

OpenAI has won major concessions from its largest shareholder, Microsoft, that will allow it to sell products on AWS, while Microsoft get more cash in a revenue-share agreement.

SpecRLBench evaluates generalization of LTL-guided RL across unseen specs and environments in navigation/manipulation.

Google Meet speech translation feature rolling out to mobile with real-time multilingual support across 6 languages and voice synthesis.

Informal benchmark comparing GPT 5.4 and GPT 5.5 on MineBench shows marginal improvements, with 5.5 Pro/standard gap narrower than prior generations.

BiLSTM and AutoML sentiment/emotion classification on Indonesian e-commerce reviews with mixed linguistic markers.

Reddit discussion comparing Claude 4.6 vs 4.7 thinking modes (extended vs adaptive) with reference to prior posts.

I’ve been playing with a Divoom MiniToo and ended up reverse-engineering enough of its Bluetooth protocol to use it as a physical Claude Code status indicator. Pretty much vibe reverse-engineering: I gave the Opus model the Android APK file to study, gently encouraged it to use jadx, and then kept letting it know whenever it crashed the device. The repo, if you're interested: [https://github.com/bugzmanov/divoom-minitoo](https://github.com/bugzmanov/divoom-minitoo) macOS only for now, but porting to other OSes should be easy The repo also includes protocol notes, a macOS Bluetooth daemon...

Learned rotary positional encodings (RoPE) as second dimension of Transformer attention expressivity.

Ineffable Intelligence, a British AI lab founded a mere few months ago by former DeepMind researcher David Silver, has raised $1.1 billion in funding at a valuation of $5.1 billion.

HyLo converts pretrained Transformers to hybrid linear-Transformer models while extending long-context via upcycling.

Case-specific rubrics for clinical AI evaluation; validates LLM-generated rubrics vs. clinician agreement on 823 encounters.

HDET explores learning rate schedules during large-model training via divergent ensembles with near-zero overhead.

Learning-based controller exploiting differential flatness for constrained multi-input robotic systems with improved computational efficiency.

Energy-Arena: dynamic benchmarking platform for operational energy forecasting with continuously updated evaluation datasets.

SpecValidator detects defective task descriptions in LLM code generation via lightweight parameter-efficient classifier.

Green Shielding: user-centric evaluation framework measuring LLM sensitivity to benign input variations via CUE criteria.

Persona Collapse in multi-agent LLM simulations: agents converge to homogeneous behavior despite distinct profiles; framework measures Coverage, Uniformity, Complexity.

SciCrafter: Minecraft benchmark evaluating agents' discovery-to-application loop via parameterized redstone circuit tasks.

Contextual Linear Activation Steering (CLAS) dynamically adapts steering strength per token, outperforming fixed-strength methods across 11 benchmarks.

Noise-Based Spectral Embedding (NBSE): physics-informed feature selection via Nishimori temperature without greedy search.