Skip to main content
Back

ARC Prize AGI-3 benchmark: Top AI models score under 1%

Cursor enables self-hosted agents, runs inside your network, Google’s Lyria 3 update adds 3-min music generation.

March 26, 2026 4 min read

Hey,

Your daily briefing is ready. You can finally take a break from the AI firehose.

Our algos spent the night splitting signal from noise and pulled the top news, models, papers, and repos.

Here’s the must-read:

Together with:
Sponsor Logo
Summary

Read time: 4 min 41 sec

Top News

▸ ARC Prize Foundation develops an interactive benchmark for agent learning

Oracle

▸ Fix stateless agents by implementing durable memory beyond context windows

Top News

▸ Google launches Lyria 3 Pro for structured 3-minute music generation

Bright Data

▸ Let AI agents browse and extract data from any URL using one CLI

Top News

▸ Cursor adds self-hosted agents that run inside your infrastructure

Signals

▸ Anthropic brings work tools like Figma, Canva, Amplitude to mobile
▸ Lambda boosts GPU utilization above 60% with reproducible training framework
▸ Google announces Vibe Coding XR for physics-aware WebXR apps
▸ Hugging Face releases hf-mount for remote storage access on local machines
▸ Anthropic shares how Claude Code auto mode automates permission decisions
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Top News
ARC Prize Foundation releases ARC-AGI-3, an interactive benchmark where agents learn rules and goals through exploration
3,484 Likes
Grok 4 Fast Benchmark

ARC Prize Foundation released ARC-AGI-3, an interactive benchmark that measures how agents learn, not what they recall.

It places agents in unfamiliar, game-like environments with no instructions, so they must figure out rules, goals, and actions through trial and error.

Humans solve all tasks immediately, while top AI systems score under 1%, showing a gap in learning ability.

What it does

ARC-AGI-3 evaluates real-time learning inside environments instead of static problem solving.

  • Uses 135 new environments designed to prevent memorization or dataset leakage
  • Requires agents to explore, form hypotheses, and update strategies during execution
  • Tests long-horizon planning where actions affect future outcomes over multiple steps
  • Measures how quickly systems acquire useful skills from interaction, not training

What it reveals

Results show consistent limitations in current agent behavior.

  • Models misinterpret tasks and apply patterns from unrelated prior experiences
  • Systems fail to revise early assumptions when new evidence contradicts them
  • Agents struggle to predict outcomes of action sequences before executing them
  • Performance remains near zero despite strong results on other benchmarks
Try arc-agi-3
Presented by Oracle
Solving the “Illusion of Memory” in AI Agents
alpha_signal_image_1

Building production AI agents at enterprise scale is fundamentally a database problem.

The “dealbreaker” in current designs is agents that forget users once a session ends. Relying on large context windows creates an “illusion of memory” that degrades before filling.

This technical deep dive explores:

  • Persistent, evolving state with vectors and graphs

  • ACID transactions in a converged engine

  • Code snippets & a GitHub notebook to get started

Explore agent memory
partner with us
Top News
Google adds section-level control in Lyria 3 Pro for coherent full-track music generation workflows
3,368 Likes
Grok 4 Fast Benchmark

Lyria 3 Pro extends Google’s Lyria 3 music model to support full-track generation with explicit structure control. It generates tracks up to 3 minutes with prompts that define intros, verses, choruses, and bridges. 

It solves the need to stitch multiple generations by letting you define song structure directly. 

KeyFeatures

  • Generate up to 3-minute tracks with consistent structure across all sections

  • Prompt song layout explicitly, like intro → verse → chorus → bridge

  • Maintain continuity across transitions instead of producing disconnected audio segments

  • Accept text prompts that describe style, structure, and musical elements

How to Use

  • Call Gemini API with structured prompts to generate audio outputs

  • Deploy via Vertex AI for scalable production workloads

  • Prototype inside Google AI Studio before integrating into your stack

  • Attach generated music in tools like Google Vids or backend pipelines

Generate tracks
Presented by Bright Data
The CLI That Lets AI Agents Browse Anything
alpha_signal_image_2

Open-source CLI with anti-bot bypass, CAPTCHA solving, and JS rendering built in.

One command returns clean markdown, JSON, or screenshots.

40+ data pipelines for LinkedIn, Reddit, X & more - plus skills for Claude Code & Cursor.

Star on github
partner with us
Top News
Cursor now lets you run cloud agents inside your own infrastructure instead of Cursor-managed environments
1,727 Likes
Grok 4 Fast Benchmark

This update targets teams that cannot move code, secrets, or build artifacts outside internal networks due to security rules. You still use Cursor’s agent system, but execution happens on your machines, not theirs.

Agents act like remote developers with their own isolated environments. Each agent can clone your repo, install dependencies, write code, run tests, and push changes for review.

Cursor handles planning and model inference, while your local workers execute the steps.

Key features

  • Runs agents inside your network; code, secrets, and artifacts never leave infrastructure
  • Uses outbound HTTPS only; no inbound ports, VPNs, or firewall changes required
  • Supports parallel agents with isolated environments for independent task execution
  • Provides Kubernetes support, worker pools, and autoscaling via deployment configs
  • Includes plugins, subagents, and multi-model support within the same agent system
Host agents local
Signals
1 Anthropic adds mobile access to Figma, Canva, and Amplitude for working directly from your phone 17,385 Likes
2 Lambda releases reproducible framework that cuts training costs by 25% or more by addressing the root causes of wasted compute Presented by Lambda
3 Google unveils Vibe Coding XR to generate physics-aware interactive XR apps from prompts 4,281 Likes
4 Hugging Face introduces hf-mount to attach Hub storage, models, and datasets as a local filesystem 1,253 Likes
5 Anthropic explains how Claude Code auto mode uses classifiers to replace frequent manual permission approvals 2,275 Likes
At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.

Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.
Looking to promote your company, product, service, or event to 250,000+ AI developers?
Work with us