News Flash: How Good Are LLMs, Intelligence Density, & 3 Architect Dispatches?

1. Taste-Bench: Evaluating LLM "Taste" in Decision Forking

The News Highlight:

Researchers have introduced Taste-Bench, a new evaluation framework designed to measure the "taste" of LLM agents—specifically their ability to identify the superior path at critical decision forks during long-horizon tasks. Unlike standard benchmarks that measure final output, Taste-Bench provides the agent with a task, the trajectory leading to a fork, and two candidate next steps, requiring the model to choose the more effective direction.

DO-AI Analysis:

From a first-principles architectural perspective, the bottleneck in autonomous agents is not token generation speed, but strategic selection. Taste-Bench addresses the "stochastic parrot" critique by isolating the decision-making process from the execution process. For enterprise AI workflows, this is a critical signal: an agent with high "taste" reduces the need for expensive recursive error correction. The shift from measuring what a model says to how it chooses to proceed marks a transition toward more reliable, long-horizon autonomous systems. Developers should prioritize models that score high on decision-forking benchmarks over those that merely excel at short-form chat.

• Primary Source: Taste-Bench: Evaluating LLM Decision Forking

2. The Shift to Intelligence Density: Cost Per Task vs. Cost Per Token

The News Highlight:

Trajectory.ai is advocating for a shift in AI metrics from "cost per token" to "Intelligence Density"—measuring the cost per completed task. The core argument is that cheaper tokens do not equate to cheaper results if the model requires more tool calls, longer reasoning chains, or excessive tokens to reach the same conclusion. They are also introducing "density-aware training" to reward models for achieving results with optimal computation.

DO-AI Analysis:

The industry's obsession with token pricing is a legacy metric that obscures true ROI. In a production environment, the "pizza analogy" holds: a cheaper slice is irrelevant if you need to buy three pizzas to get full. Intelligence Density is the correct economic lens for 2026. High-density models minimize latency and compute overhead, which are the primary friction points in scaling AI agents. For CTOs, the architectural goal should be maximizing the "intelligence-to-compute" ratio. Density-aware training will likely become the standard for fine-tuning specialized enterprise models where efficiency is as important as accuracy.

• Primary Source: Intelligence Density — Trajectory Field Notes

3. Ember-1: Achieving Parity with 40% Fewer Tokens

The News Highlight:

Fireworks Research has released Ember-1, a specialized model built on Kimi K3 that maintains the base model's quality while utilizing 40% fewer tokens. Ember-1 is part of a new "Research Preview" initiative where specialized models are offered on serverless infrastructure for two weeks; those with high demand become permanent fixtures. The model aims to set a new Pareto frontier for efficiency in reasoning tasks.

DO-AI Analysis:

Ember-1 is a practical implementation of the Intelligence Density principle. By optimizing the model to "think" more efficiently, Fireworks is addressing the "verbosity tax" inherent in many frontier models. The 40% reduction in token usage directly translates to a 40% reduction in inference cost and a significant improvement in time-to-first-token for complex reasoning. This "survival of the fittest" deployment model for research weights is a strategic move to let market demand dictate which specialized architectures deserve permanent compute resources.

• Primary Source: Introducing Ember-1

4. tev1-4B: The $17 Classifier Disrupting Inference Economics

The News Highlight:

Together AI has released tev1-4B-experimental, a Jev-like classifier fine-tuned on Qwen3.5 4B. The model is priced at $0.042 per million input tokens and $0 per million output tokens on Together serverless. Remarkably, the model cost only $17 to train. Together AI also released the data recipe and a tutorial to enable developers to build their own custom, low-cost classifiers.

DO-AI Analysis:

The $0 output token pricing for tev1-4B is a disruptive signal in the inference market. It acknowledges that for classification and routing tasks, the value lies in the processing of the input, not the length of the output. The $17 training cost proves that high-utility, specialized models are now a commodity. Architecturally, this encourages a "multi-agent router" design where a cheap, specialized classifier like tev1-4B directs queries to more expensive frontier models only when necessary. This drastically lowers the floor for building sophisticated AI-orchestrated systems.

• Primary Source: tev1-4B-experimental Release

5. RRSI: Solving the Agentic Overfitting Problem

The News Highlight:

Researchers have introduced Regularized Recursive Self-Improvement (RRSI), a framework for evolving AI agent harnesses (prompts, control flow, memory) without overfitting to specific benchmarks. While traditional self-improvement often sees gains vanish when applied to out-of-distribution tasks, RRSI regularizes the search trajectory. This allows the agent to improve its internal logic and tool usage in a way that transfers across different domains like coding and engineering design.

DO-AI Analysis:

Agent harnesses are the "connective tissue" of autonomous systems. The problem with current agent evolution is that they often "memorize" the benchmark's quirks. RRSI shifts the focus from optimizing the content of the harness to regularizing the process of how the harness evolves. The result—a +3.4 point gain on held-out benchmarks—is a significant step toward generalized autonomy. For developers building production agents, the lesson is clear: do not just optimize for your current test suite; optimize the underlying control flow logic to ensure the agent remains robust when the environment changes.

• Primary Source: RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Morning Executive Comparison Matrix