Reading Room

Cheap, fast decision models like this one could end up doing a lot of the small routing and classification work inside your agents (AI systems that take several steps on their own), instead of calling a big, expensive AI model for every tiny choice. This week the team behind the Strands agent-building tools released Strands Decider 2B, a small model built to do one narrow job: pick between a short list of options, or score something, rather than write open-ended text the way a normal AI model does.

These are sometimes called decision models or system one models, a category that has been getting attention since a company called TypeSafe AI launched its own version, named Jev, earlier this month. A normal AI model (a large language model, or LLM) can write anything you ask it to. A decision model cannot. It only answers questions like “Is the phrase ‘turn on the lights’ about the coffee machine, yes or no?” or “What language is ‘sihamba ngokushesha’: English, Zulu, or Dutch?”, or gives a score, such as rating how positive a sentence sounds on a scale of zero to one. In exchange for that narrower job, these models are faster, always give one of the listed answers, and run with very little delay. The trade-off: they are much worse than reasoning models (AI that works through a problem step by step) at hard, multi-step problems, and they cannot write text at all, so they are useless for coding, chatbots, or summarizing documents. What they add that a normal AI model does not is a reliability score for each decision, how sure you can be the answer is right, and they are cheap enough to ask many questions about the same piece of text at once.

Strands Decider 2B has 2 billion internal settings (parameters), small enough to run on a laptop chip or a single graphics card you already own. Under the hood, the team took an existing AI model (called Qwen3.5-2B) and removed its text-writing part, replacing it with a much smaller scoring component whose only job is to score each possible answer against the question. The main model underneath got a light round of extra training using a method called a rank-16 LoRA adapter, a cheap way to retrain a small slice of a model instead of the whole thing. The team says this is the model's 19th version; an earlier design, which scored each answer differently, performed worse and was dropped.

On the field's standard test, JevBench (a standard test everyone in this category runs), Strands Decider 2B ranked 3rd out of 33 similarly sized models on accuracy, and on calibration, how trustworthy its confidence scores are, measured using something called a Brier score, it ranked 1st out of 30 once models just over 2 billion settings are excluded. Both scores improved steadily across the 19 versions. It answers in a median of about 153 milliseconds on an Apple M3 laptop chip for small questions, with similar speed on the Nvidia RTX 3090 used to measure it; answers take a little longer as the question itself grows. The team picked 2 billion settings on purpose: small enough that people can run, and even retrain, it on hardware they already own, big enough that it gets 100 percent of the easy questions on JevBench right. It released the model, its training data, and the scripts used to build it for free on GitHub and Hugging Face (two sites that host AI code and model files).

This is on your desk because two companies released the same kind of specialized AI model in the same week, a sign this is becoming a real category, worth knowing if you design AI agent systems that need cheap, fast decision steps.

Microsoft this week introduced Microsoft-Decision-1, built specifically for fast decision-scoring: quick yes or no calls, picking from a list, or rating something. Microsoft draws a clear line between this and large language models (AI models built to write open-ended text or reason through a problem): a decision model is built only to hand back a structured answer that software can act on immediately, nothing more. Microsoft says that narrower job is what lets it run decisions and sort things at very low cost with high performance, which is why it says many useful tasks become possible once you have a model built this way.

The model is available in Microsoft Foundry, Microsoft's AI model hosting service, and through OpenRouter, a marketplace that lets you call many different AI models through one connection. Microsoft built it for routing requests, classifying text, setting priority, double-checking other AI output, and controlling the steps in automated workflows, aiming to make it easier to add this kind of decision-making to existing apps, agents, and workflows inside a secure, trusted setup.

On the numbers: Microsoft says it tested the model against 36 standard tests covering close to 150,000 questions kept hidden from its training data, and that it came out the most accurate of the models it compared against. On speed, Microsoft says it was the fastest in the test: 4.5 times quicker than the runner-up, a model called Quyet-1.0-Large, and 35 times quicker than a model called GPT-6 Sol.

This lands in the same week as Strands Decider 2B, a smaller model built on the same idea by the Strands agent-building team, a sign that more than one company now sees small, specialized decision models as worth racing to ship.

Outside your usual reading: this is a sharp, real example of what happens once surveillance tools get cheap and easy to build. It can cut both ways, once ordinary people can build the same gear governments use.

Police departments have been installing automatic license plate readers, cameras known as ALPR (they scan car license plates and log where and when each car passes), often sold under the brand name Flock. The technology is not new, but its use has exploded over the past few years, and it has been tied to real misuse: stalking ex-partners, harassing immigrants, and wrongly flagging random people. Canadian software engineer and YouTuber Anthony Sistilli, from Brampton, Ontario, a city that recently spent 2 million Canadian dollars on its own high-resolution ALPR system, decided to build his own version and point it only at police cars.

Sistilli posted TikTok videos explaining how he built the camera, describing it as a personal Flock surveillance camera to Flock the Flockers. A month or so later, two officers from Peel Regional Police showed up at his front door to ask about his intentions for the data he was collecting. He said one officer told him it would be scary for police if people could find out where they lived, when their shifts started, and what they were doing throughout the day, since that kind of detail could put them at risk from any bad apples who might want to track them down.

Sistilli used the visit as the opening of a longer YouTube video that looks at how Flock markets its cameras to local governments in what he calls an aggressive and arguably deceptive way, whether the cameras actually cut crime, and how little is disclosed about how long the footage and location data get kept. Peel Regional Police did not respond to Gizmodo’s request for comment.

A new Reuters and Ipsos poll finds most US voters do not think President Trump or Congress are taking the risks of AI seriously. It is one data point in the slow build toward AI rules that will eventually affect anyone building AI products, so it is worth tracking even at this headline level.

A report argues Meta's push into AI tools for businesses could turn it from just another AI feature vendor into a direct rival to the subscription software, called SaaS, that companies already pay for. If big platforms start competing with the tools your business already uses, that changes who you build and buy software from.

Anthropic updated its usage policy, banning sustained, needless cruelty toward Claude alongside new limits on propaganda campaigns, surveillance, and weapons development. Ending the conversation stays the main penalty, reserved for extreme, repeated cases, not ordinary pushback or testing, which gives a window into how the company shapes what you are allowed to do with the model you use every day.