PixelsTech feeds

Stay current. Skip the noise.

The right amount of technology news each day, curated for developers.

Welcome to PixelsTech Feed — daily tech news curated by AI across topics like Go, Python, AI, Rust and more.  How it works →   Log in for full access →
Subscribe via RSS

2026-11-15

Alibaba's Qwen4-Max beats Qwen3.8-Max with half the tokens and a 62% lower cost per task
Alibaba's Qwen4-Max uses up to 50% fewer tokens than Qwen3.8-Max and cuts average task cost 62%, with native vision and a 1M-token context window.
The efficient frontier of LLM inference
Inference techniques either move a deployment along the latency–throughput frontier or push the entire frontier out, creating more efficiency to allocate.
Eval Awareness vs. Eval Paranoia
I have mentioned several times that I think frontier models are not that eval aware. I think this is an important fact to understand about current mo…
GPU says 35% busy. Busy doing what?
nvidia-smi said our H100 was 35% busy while users queued for it. Counting every kernel and memory copy found the cause, 643 MB copied off the GPU per request.
Understanding Positional Embeddings in Modern LLMs
The Methods Behind How LLMs Represent Token Positions
What thousands of conversations look like inside a language model
I mapped thousands of conversations inside Qwen to see how language takes shape.

2026-11-14

What a task costs on Opus 5.5
The same token price can cost very different amounts per task. Learn what Claude Code tasks cost on Opus 5.5 and which settings change the bill.
Gemini 4 Carbon tops MathArena but costs far more than GPT-6.1 Sol
Google's Gemini 4 Carbon (high) now ranks first overall on MathArena with an expected performance of 85.4%, ahead of Anthropic's Claude Fable 5.2 (max) at 82.8% and OpenAI's GPT-6.1 Sol (max) at 80.6%. MathArena added both new models' results on November 13. The Decoder reports that Carbon and Fable 5.2 cost roughly 15 and 40 times as much as GPT-6.1 Sol under the reasoning settings used in testing.
Under the Hood: Serving Kimi K3 | DigitalOcean
How DigitalOcean served Kimi K3 on day 0: GPU selection (B300/MI350X), vLLM tuning, and the OpenAI-compatibility fixes for tool calling, streaming, and reasoning.
Serving GLM-5.3 at 607.8 tok/s across two Nvidia DGX stations with vLLM
Running a frontier-scale model locally means keeping all its weights in GPU memory; the largest models no longer fit on one machine. Two NVIDIA DGX Stations…
Generative AI uses more energy on non-English text
When processing and producing non-English text, generative AI burns more energy and sometimes provides less accurate answers, potentially reinforcing existing global inequalities, according to University of Michigan Engineering researchers.