"Tokenmaxxing" Spreads Inside Amazon — AI Usage Leaderboard Manipulation Becomes a Problem

It has come to light that "Tokenmaxxing" — a practice in which Amazon employees inflate their AI tool usage — is spreading within the company. Employees reportedly automate unnecessary tasks to deliberately burn tokens, all in pursuit of ranking higher on internal AI leaderboards.

Amazon has introduced automated prompt optimization in its Bedrock AI service, delivering performance improvements of up to 22% depending on the task. Anthropic and OpenAI offer similar optimization tools, but accurately evaluating optimization results itself remains a challenge for the industry as a whole.

The issue highlights a side effect of companies measuring AI adoption by usage volume. Poorly designed metrics risk pulling organizations away from AI's original purpose: improving productivity.

Statewright — Tackling AI Agent Reliability at the Root with State Machines

Ben Cochran, a Distinguished Engineer who worked at NVIDIA and AMD, has announced "Statewright," which takes on the AI agent reliability problem with a state machine approach.

He points out that current agent approaches rely on brute force — either making the model bigger or widening the context window. Instead, he adopted an approach that shrinks the problem space. Formal state machines constrain models in the 13–20B parameter class, defining which tools are available in each state, how many iterations are allowed, and the transition conditions.

The results exceeded expectations, with consistent improvements across multiple model families. Haiku and Sonnet reportedly perform above their class, while Opus can now solve problems more reliably with fewer tokens. The key insight: how efficiently the context window is used matters more than its raw size.

Statewright is built on a Rust engine core and integrates with Claude Code as an MCP plugin, with support for Codex, Cursor, and others planned down the road.

TabPFN-3 Released — A Foundation Model Supporting 1M Rows of Tabular Data

TabPFN-3, a pretrained foundation model for tabular data, has been released. Traditional tabular data approaches spent considerable time on feature engineering and hyperparameter tuning, but TabPFN dramatically reduces that effort through neural network-based pretraining.

The biggest improvement is that supported dataset sizes have been expanded to up to 1 million rows. Since much of the practical enterprise data out there falls within this scale, competition with established tools like XGBoost and LightGBM is becoming a realistic prospect.

Dessn Raises $6M for an AI Design Tool Wired Directly into Production

Startup Dessn has raised $6 million to develop an AI design tool that works directly with production codebases. While traditional design tools stop at static mockups, Dessn proposes design changes based on the actual production code.

This approach of cutting the back-and-forth cost between design and implementation is worth noting from a developer workflow efficiency perspective.

Kimiflare — An Open-Source Coding Agent Running Kimi K2.6 on Cloudflare Workers AI

Prompted by the expiration of Cloudflare credits, an open-source CLI coding agent called "Kimiflare" has been built with Kimi K2.6 as its backend. Positioned as a Claude Code clone of sorts, it can be installed via npm.

It has racked up more than 12,000 npm downloads and 137 GitHub stars. Particularly interesting is the dogfooding in practice: the first few commits were made with Claude Code, and from v0.3.0 onward, development has continued using Kimiflare itself. After the Cloudflare credits expire, it plans to transition to a BYOK (Bring Your Own Key) model.

Big Tech's AI Investment Debt Surges — Spending Set to Expand Further

According to Barron's, AI-related investment debt at big tech companies is surging, with spending projected to expand even further. Massive investments in AI infrastructure are outpacing the speed of monetization, and debt levels are becoming a source of concern.

At the same time, the competitive risk of underinvesting in AI is significant, leaving companies facing a dilemma: invest or lose. This dynamic looks set to become a major theme for the tech sector over the next several quarters.