An agent does not pay for memory once. It pays again every turn
Storing a memory is only half the problem. On every model call, an agent still has to decide what deserves context space, what can be compacted and what should stay outside.
Archive
AI
IRZ news, analysis and ideas about AI.
CAD agents need work traces more than perfect demos
CADBench, Task Model Induction and Skill Recorder point to the same constraint from different angles: automating expert software requires verifiable traces of real work, not just a polished demo.
WebMCP won't replace scraping. It replaces blind clicking
WebMCP lets sites declare actions directly to browser agents. It is an alternative to brittle DOM actuation, not the end of scraping or dedicated APIs.
The AI model race is starting to measure everything around the model
In fifteen days, Qwen 3.8, DeepSeek V4 Pro, Muse Code, Grok 4.6, Gemini 3.7 Flash and several open models arrived. The more interesting change sits around them: the harness is becoming a measurable part of the product.
How Midjourney was built: open diffusion, 10,000 GPUs and zero VC
Midjourney began as a small self-funded lab building on open breakthroughs in CLIP and diffusion. What we know about its first models, training costs and shift toward proprietary technology.
Agent rules accumulate because we forget why they were added
A study of 1,867 repositories finds a ratchet effect in CLAUDE.md-style instruction files. Adding a rule is cheap. Removing it requires remembering why it exists.
The same words can change a vision model’s decision when typography changes
The same words can produce a different judgment when a model sees them as an image. ThinkingType turns presentation drift into a reproducible benchmark.
Claude is adding invisible marks to its output, but they do not prove who wrote it
Anthropic is starting to mark Claude output with invisible text watermarks and signed provenance for supported files. The useful signal is “processed by Claude,” not “authored by Claude.”
Local AI has a new problem: turning a GPU into a machine people can simply use
Cloudless aims to move beyond DGX Spark onto ordinary GPU computers running Linux and Windows through WSL, without reformatting them. The detail points at local AI’s next bottleneck: making installation disappear.
116 ways an LLM benchmark can be wrong without an obvious error
The community registry tracks 116 serving-path traps where requests look valid, responses look valid, and the final measurement is still wrong because of templates, parsers, kernels, runtimes or version drift.