Latest in AI
Every brief, newest first. Each one links to its primary source.
-
Anthropic’s September misuse report names specific operations
Anthropic’s September report names specific operations (twenty-plus government targets, 1.8 million Android APKs, 8,913 planted articles), all of it vendor-detected and unverifiable from outside.
-
Injected plans can slip past chain-of-thought monitors, researchers report
A new preprint reports 25 to 33 per cent monitor evasion by hiding a harmful plan in a model’s context, and finds that showing…
-
Nvidia adds a fault-tolerance layer to its open-source quantum stack
Nvidia’s new CUDA-Q Logical layer targets quantum error correction rather than qubit counts, though the eye-catching speed-up figures come from partners describing their own…
-
How Self-Driving Actually Works: SAE Levels 0–5 and the State of Play
The SAE levels are the only precise vocabulary for autonomous driving, and almost every consumer claim abuses them. Here is the framework and what…
-
The AI Compute Stack: Why Chips Decide the Race
AI capability is downstream of hardware. Understanding the compute stack (accelerators, memory, packaging, power) explains most of the industry’s strategic behaviour.
-
AI Benchmarks Explained: What the Scores Actually Measure
Benchmark numbers drive coverage, procurement and valuations. Most of the widely-quoted ones are saturated, contaminated, or not comparable between models.
-
Why AI Models Hallucinate
Hallucination is not a bug in language models. It is a predictable consequence of how they are trained and, crucially, of how they are…
-
Tool Use and MCP: How AI Connects to the Real World
Tool use is how a model stops merely talking and starts doing. MCP is the standard that stopped every integration from being bespoke.
-
Fine-Tuning vs RAG vs Prompting: Choosing the Right Approach
Three ways to make a model do what you want, routinely confused with each other. The deciding question is whether your problem is knowledge,…
-
Reasoning Models Explained: Chain-of-Thought and Test-Time Compute
Reasoning models spend extra computation thinking before answering. The gains on hard problems are real, the costs are substantial, and the tradeoffs are not…