@alphasignalai

–

27 May 2026 8h55

input tokens are the real driver here. agent re-reads the entire context on every single loop

27 May 2026

MIT Study: AI Agent Token Costs Dominate Inference Economics

By

–

27 May 2026 8h21

MIT just killed the idea that cheaper tokens mean cheaper AI. A paper from MIT, Stanford, Google DeepMind, and Microsoft found agents use 1000x more tokens than chat. The cost isn't the code they write. It's re-reading context on every loop. Same task, same model: 30x cost

27 May 2026

AI Agents Transform Codebases into Knowledge Graphs

By

–

26 May 2026 17h38

You probably still understand a codebase by reading files manually. This open-source repo turns entire codebases into interactive knowledge graphs using 7 parallel agents. It maps business logic alongside architecture and lets you trace dependencies, flows, and systems visually

AI Agent Token Consumption: Research Paper & Project

By

–

26 May 2026 17h28

Paper: https://
arxiv.org/abs/2604.22750
Project website: https://
longjubai.github.io/agent_token_co
nsumption/
…

Researchers Identify Neurons Behind AI Safety Refusals

By

–

26 May 2026 14h01

Someone just found the exact neurons that make AI say "no." Language models refuse harmful prompts, but nobody knows how that refusal works inside. Most steering methods edit the residual stream and wreck output quality. A new paper proposes a sharper fix: Contrastive Neuron

Training Agents with Failures as Optimization Signal

By

–

26 May 2026 13h45

The trick in our case is treating failures as training signal, which only works with a reliable verifier and a held-out gate. Open-ended agent work without success metrics is the hardest case to optimize this way.

Multi-Turn Agent Evaluation and Context Compaction Limits

By

–

26 May 2026 13h43

Paper does test multi-turn (24 tool calls in OfficeQA, 30 turns in SpreadsheetBench, 50 steps in ALFWorld), but mid-session auto-compaction isn't part of the eval. Skill being under 2K tokens probably helps it survive compaction, but not validated against that failure mode.