A cheap judge gives only the odds for each verdict, and its confidence decides whether to accept it or hand the task to a slower reasoning judge. On 1,610 held-out pairs this routing beat GPT-6 by 0.9 points at 41% of its cost.

Loom is a compiler that turns chip layout and data movement into a puzzle solved at compile time, picking placement, communication, and tile sizes instead of hand-tuned kernels. On two chip generations it matched or beat the vendor's kernels on three AI operations without tuning.

After each coding attempt, the agent writes an explanation of what worked or failed, and is trained only on those explanation words—no rewards or outside teacher. It solved 49.2% of held-out coding tasks versus 48.0% for reward-based training, using half the updates.

Tagging one of two meaningless zones with a 'good' or 'bad' internal brain pattern, then switching it off, showed which zone the model preferred. Models picked the good zone even when every visible word was identical, and worked to erase bad states.

Instead of pasting a fixed copy of every file it reads into its history, the agent keeps a live list of files and pulls in their current contents at each step. This cut input tokens by 9-50% while solving the same share of tasks.

A chatbot saves the math from reading a prompt so follow-ups come fast; today each cloud keeps its copy. This vision shares those saved chunks across clouds, deciding where to store or recompute them by network speed and price to cut delay and cost.

A model teaches itself using a copy that sees the answer as a guide, and both copies keep improving each round so new gains feed back in. On an 8-billion-parameter model this hit 65.97% on math contests, beating the frozen-guide version by 35.62 points.

A memory system that turns each new experience into a note with keywords and tags, then links it to related past notes and updates those notes as understanding grows. Tested on six different AI models, it beat the best existing memory systems at helping agents use past experience.

Fully homomorphic encryption lets a server run a model on scrambled data it never sees; this system repacks data so weights and results need less re-encoding. It runs Llama 3 8B on encrypted input on one GPU, 4.51 times faster than the earlier approach.

A study counted the distinct real-world claims 27 chatbots produced across 155 topics in 12 countries over three years, comparing them with web search results. Variety has grown over time but still trails search, and English knowledge crowds out local-language knowledge.