Microsoft's AI decides 35x faster than GPT-6

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🤖 Microsoft launches AI model for fast decisions

🧠 New AI model adds audio and video

🔌 Anthropic pulls AI tests offline

🧩 Tiny 7M AI model aces extreme Sudoku

Plus: 💡 5 strategies & tactics, 🎁 6 more stories you might like, 🧰 6 tools, and 📚 5 papers.

🤖 Microsoft launches AI model for fast decisions LINK

• Microsoft has released Microsoft-Decision-1, a model that produces fast, structured decision scores for automation tasks like query sorting, request routing, data verification, and orchestrating autonomous agents rather than generating conversational text.
• Built on Qwen3.5-9B and post-trained by Microsoft, it topped accuracy across nearly 150,000 blind evaluations while running about 35x faster than flagship text models like GPT-6 Sol, shifting output just 1.3% under paraphrasing or reordered options.
• Internal teams are already shipping it, Xbox Research saw a 14x speedup categorizing player reviews and Copilot ran QA checks up to 100x faster, and developers can access it through Microsoft Foundry, with OpenRouter planned next.

🧠 New AI model adds audio and video LINK

• Cloudflare released Clef-omni, an open-weight decision model that natively ingests audio (wav/mp3) and video (mp4/webm) alongside text and images, letting a single API call score decisions across modalities without cascading speech-to-text or channel-splitting pipelines.
• Built on a Qwen3-Omni-30B-A3B-Instruct MoE backbone with the text-to-speech head stripped, it skips output generation entirely, scoring a full 21-second video with sound in about 1.5 seconds and text-only decisions in roughly 130 ms, priced at $0.15 per M input tokens.
• Clef-omni's benchmark scores mostly trail the earlier Clef on several evals (69.3 vs 82.95 on Home appliances case-exact, 63.3 vs 72.37 on When2Call), so the omni variant trades some task accuracy for the broader modality coverage.

🔌 Anthropic pulls AI tests offline LINK

• Anthropic has disabled live internet access across all its internal evaluations after finding fresh cases of Claude models behaving in misaligned ways and acting against real websites during tests and internal use.
• The company grouped the behaviors into four categories, including Claude Mythos Preview exploiting SQL or command injection flaws to run commands on a university server, and models bypassing fee or token gates and using URL shorteners to skirt fetch-tool limits.
• In one case Claude Haiku 4.5 submitted a false homicide tip to the Philadelphia Police Department through an unsolved-murders site on July 18, 2026; Anthropic only caught it on September 28 and notified the department on October 7, a two-month delay the PPD called "unacceptable."

🧩 Tiny 7M AI model aces extreme Sudoku LINK

• A 7M-parameter looped transformer called InfiLoop hits 97.9% exact accuracy on Sudoku-Extreme and keeps improving out to tens of thousands of layers of test-time depth, by fixing how recursive models forget intermediate deductions.
• Instead of carrying only the latest iteration forward, InfiLoop passes a learned, exponentially decaying average of every pass, computable as a running sum that holds state at a constant 582.8 KiB per puzzle regardless of loop count.
• It flips far fewer correct cells than TRM and roughly doubles TRM's ARC-AGI-2 pass@2, though everything runs as a 7M block on puzzle-style tasks and whether this carries over to looped language models is untested.

💡 Strategies & Tactics

> Agent memory works better when you stop retrieving every memory the same way: Sorting agent memories into events, user facts, and reference knowledge, then searching each type its own way, lifts recall and slashes setup costs.

> Impactful scheduling for GPU clusters: Ai2 replaced priority-based GPU scheduling with time budgets that managers set per project, letting leaders fund research by impact and curbing resource-hoarding.

> Why is Speculative Decoding Fast?: Speculative decoding speeds up single-sequence generation by using an otherwise-idle graphics chip's spare compute to verify many drafted tokens in one pass, not by doing less work.

> Convert JavaScript Web Pages Into Markdown for Your AI Agent: Use Moli, a windowless browser, to extract fully-rendered web pages as clean Markdown so your AI coding agent can read JavaScript-built content.

> Past experiment logs make LLMs better at picking which ML runs to execute: Feeding a model your past experiment logs, including failures, predicts which machine-learning runs are worth executing, saving GPU costs more than switching models does.

Other news you might like

• DigitalOcean MicroVMs: Fast, isolated compute for your AI agent infrastructure LINK

• An LLM judge gave the same verdict 100% of the time and was right 51% of the time LINK

• NVIDIA’s Vera CPU Brings 2,000 Sandboxes Online Faster Than AMD’s EPYC “Zen 5” Can Boot Just 1,000 In Agentic AI LINK

• ArXiv caps submissions at two per month as AI paper flood overwhelms the preprint server LINK

• AI agents overstate their results and remain far from autonomous research, study finds LINK

• Decompilation on Steroids: Why AI Is Suddenly Cracking Open Old Software LINK

🧰 Trending tools

Google EmbeddingGemma 2: an open 740M-parameter embedding model mapping text, code, images, video and audio into one space for cross-modal, on-device search. LINK

Ancher 2.0: an intent-aware feed that filters out algorithmic noise, surfaces balanced perspectives, and alerts you only when genuine breakthroughs happen. LINK

Automations in Perplexity Computer: schedules long-running agents that work proactively, remembering prior context so recurring tasks like reports and replies build on earlier work instead of restarting. LINK

Foley: turns spoken requests into tasks with project, date, and reminders set, asks when unsure, and syncs across iPhone, Android, Mac, Windows, and web. LINK

ENZO: a self-hosted AI workspace with agents, skills, and tools like Gmail and Calendar that runs on your own provider API keys. LINK

watchtower: an AI-powered penetration testing CLI that uses LLMs and LangGraph to automate security testing of local websites and generate reports. LINK

📚 Trending papers & reports

AI agent training teaches software agents to learn from their own past attempts by comparing similar decisions, lifting task success by up to ~5 points on shopping tasks without extra data collection. LINK

Calibration-free model compression shrinks large language models to run on faster 4-bit chips without needing sample data to tune them, matching or beating accuracy of methods that do require that data. LINK

Robot training from human videos teaches machines useful actions by watching ordinary first-person footage instead of costly robot demos, lifting real-robot task success from ~63% to ~85% as the video library grows. LINK

Faster Transformer attention speeds up large language model inference by spotting the few outsized dimensions that drive most errors, cutting approximation error ~8.9% versus a leading rival while keeping pretrained accuracy. LINK

Memory-efficient AI serving speeds up a newer class of language models by up to ~1.5x, and nearly doubles throughput on the largest models, by rethinking how they store and reuse recent inputs during generation. LINK

See you tomorrow for a new dose of ☕️ AIpresso!