Lab
Research, experimentation and open source: hardware, IoT, robotics, biotech and edge AI.
50 published articles
DeepSeek-V4.1-Flash lands on Cambricon's vLLM stack
A 552B model, a quarter of the HBM, and a day-one Cambricon port
DeepSeek-V4.1-Flash reached Cambricon accelerators on day one through vLLM. The port says less about the model than about how fast Chinese chip vendors can now follow a public inference framework.
2026-09-22
Camera-trap AI · Species ID benchmarks, September 2026
BioCLIP, 300M parameters, beats bigger vision models at species ID
A 300M-parameter specialist outperformed four vision-language models in the 2B-to-8B range on a 96-species task. The same study found field imagery degrades every model, and up to 9.6% of open-set answers name species that do not exist.
2026-09-21
AI research · Structured output benchmarks
LFM2.5-350M gained 14 points on JSON. YAML barely budged
A 100-step GRPO run lifted LFM2.5-350M from 22.6% to 29.7% on the IFStruct schema benchmark. JSON gained 14 points, YAML gained 0.3, and the dominant failure mode barely moved.
2026-09-17
Humanoid Robotics
Beijing's robot games were 96% Chinese. That's the real story
A 9.39-second 100-meter sprint and a 2.88-meter standing jump made the second World Humanoid Robot Games look like a robotics breakthrough. The real signal ran deeper: 641 of 666 teams were Chinese, and the biggest US humanoid makers stayed home.
2026-09-16
Mergers & Acquisitions
GoPro's $285M sale to Starman pushes its cameras toward defense work
Starman Holding is buying GoPro for $285 million in cash and steering it toward defense, government, robotics and aerospace. The deal follows years of decline, $92 million in debt, and a YouTuber's surprise 8.5 percent stake.
2026-09-15
Financial reasoning benchmark
LLMs know accounting formulas. FinIndices shows they can't apply them
LLMs can recite accounting formulas, but FinIndices, a benchmark built on real Chinese financial statements, shows they struggle to apply them. Gemini-3.1-Pro's accuracy fell from 70.70% to 38.22% when formula hints were removed, and generating tables actively degraded models' reasoning.
2026-08-20
HDR10+ ADVANCED: Samsung and Prime Video ship it first
Can scene-by-scene metadata fix motion smoothing's bad reputation?
Prime Video is the first streaming service to adopt HDR10+ ADVANCED on Samsung's 2026 TVs from August 2026. The standard adds Enhanced Overall Brightness and scene-aware Intelligent Motion Smoothing, the setting with the worst reputation in home theater, now backed by smarter metadata.
2026-08-11
AI in Education
AI tutors over-help by design. TutorMoments quantifies it
A benchmark built on 462 real tutoring sessions finds AI tutors default to doing the work for students. Explicit prompting helps, but models still struggle with the core judgment call: when to push, when to scaffold, when to step back.
2026-08-10
Voice Agents
Why Grok 4.1 Fast beats smarter models on a phone call
Overall leaderboards pick the wrong LLMs for voice calls. BenchLM's 2026 ranking puts latency first: Grok 4.1 Fast answers in 0.54s while Claude Opus 4.6, the top scorer, needs 1.78s. Fast models take the conversation; reasoning models stay on background tool calls.
2026-08-05
AI Evaluation
The old AI benchmarks broke. Here's what replaced them.
Static benchmarks like MMLU and GSM8K are saturated and contaminated. The industry has moved to dynamic regeneration, expert-level exams, and agentic task environments to get a real measure of AI capability.
2026-08-04
Architecture & Cloud
Why your IoT dashboard shows yesterday's data, and how to fix it
Most IoT monitoring platforms fail not at data collection but at reliable delivery. A layered architecture using Alibaba Cloud DataWorks automates scheduled synchronization, eliminates risks like duplication and latency, and keeps your dashboards current without custom middleware.
2026-08-03
Special Report: AI Evaluation
The State of AI Benchmarking in 2026: Inside the Collapse of Static Tests and the Systems Built to Replace Them
A comprehensive tour of the 2026 AI benchmarking landscape: why MMLU, GSM8K, and HumanEval broke, how dynamic benchmarks and expert exams like HLE and GPQA Diamond replaced them, what agentic and jagged-intelligence testing reveals, and how human preference, LLM judges, and production observability now round out the full evaluation stack.
2026-08-03