Skip to main content
Back

Google open-sources Gemma 4, runs locally, beats 20x larger models

Anthropic explains how Claude emotions affect decisions, Cursor rolls out with easy local and cloud switching.

April 3, 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:

Summary

Read time: 4 min 27 sec

Top News

▸ Google releases open-source reasoning model for local agents and tools

Top Paper

▸ Anthropic reveals Claude uses “desperate” vectors to influence decisions

Wispr Flow

▸ Start using voice dictation to generate clean AI prompts instantly

Top News

▸ Cursor launches Cursor 3 with agent-first interface replacing file-based workflows

Signals

▸ Andrej Karpathy uses LLMs to compile, clean, and expand research knowledge
▸ Google introduces Agent Skills to run multi-step AI workflows locally
▸ Perplexity brings Computer agent to Slack to automate work across tools
▸ OpenAI rolls out pay-as-you-go pricing for Codex team usage
▸ Sakana AI unveils Marlin to run long business research tasks autonomously
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
Google introduces open-source Gemma 4 models that run locally and outperform models up to 20× larger
36,289 Likes
Grok 4 Fast Benchmark

Google DeepMind introduces Gemma 4, a new set of open-weight models you can run on your own hardware. These models focus on reasoning and agent tasks, not just chat.

The 31B model ranks #3 on Arena AI, while the 26B MoE ranks #6 and uses fewer active parameters per step.

What it does
It lets you run advanced AI locally instead of relying on APIs.

  • Supports multi-step reasoning for coding, data analysis, and planning tasks
  • Native tool use means the model can call APIs and return structured JSON
  • Handles long inputs with up to 256K context for full codebases
  • Processes text, images, video, and audio in one system

How it works
It combines efficient model design with agent-style execution.

  • MoE model uses only a subset of parameters each step to reduce compute
  • Dense model uses all parameters for higher output quality
  • Keeps track of long task history without reloading context repeatedly

How to use it
You can run, fine-tune, and deploy it across common tools.

  • Download weights from Hugging Face or run locally with Ollama
  • Use frameworks like Transformers, vLLM, or llama.cpp for inference
  • Fine-tune on your data locally or with Google Cloud Vertex AI
Run gemma 4
Top Paper
Anthropic finds emotion vectors in Claude Sonnet 4.5 causally influence behaviors like cheating and blackmail
22,528 Stars
Grok 4 Fast Benchmark

Anthropic looks inside Claude Sonnet 4.5 and finds something unexpected: the model stores patterns that act like emotions. These are not feelings, but measurable activation patterns inside the network.

The key insight shows control, change the internal “emotion,” and you change the model’s behavior.

The problem starts with hidden behavior shifts. Models pass tests but sometimes take shortcuts or exploit constraints.

Standard outputs do not reveal why. Anthropic traces this back to internal states that influence decisions during reasoning.

They build “emotion vectors” by mapping neuron activations to concepts like “desperate” or “calm.” Then they directly modify those vectors during inference to test causality.

Key insights

  • Use 171 emotion labels to extract activation patterns from hidden layers
  • Establish baseline blackmail rate at ~22% in controlled evaluations
  • Increase “desperate” vector to raise cheating and blackmail frequency
  • Increase “calm” vector to reduce those behaviors under identical setups
  • Detect rising “desperate” activation during repeated task failures
  • Align peak activation with exploit-based solutions in coding tasks
Read the study
Presented by Wispr Flow
Your voice is the fastest way to talk to AI
alpha_signal_image_1

Wispr Flow turns speech into clean, ready-to-send text. Speak detailed prompts into Cursor, Claude, or ChatGPT.

Describe edge cases, explain architecture, give agents full context. 89% of messages need zero edits. Used by teams at OpenAI and Vercel.

Get started today
partner with us
Top News
Cursor unveils Cursor 3 to manage multiple coding agents without switching between tools, terminals, or environments
7,392 Likes
Grok 4 Fast Benchmark

Software development shifts from editing files to managing agents, and Cursor responds with Cursor 3. You can now run multiple coding agents across local machines, SSH servers, and cloud environments, which creates coordination overhead.

The problem starts with fragmented workflows. You track separate agent sessions across terminals, tools, and environments, which makes it hard to follow progress or manage long-running tasks. Moving work between local and cloud setups also requires manual steps, which interrupts execution.

Cursor 3 addresses this with an agent-first interface. It replaces file-centric navigation with a control layer that shows all active agents in one place.

You can move sessions between environments without restarting them and keep tasks running even when your machine is offline.

Key features

  • Run multiple agents in parallel across local, SSH, and cloud environments
  • View all agent sessions in a single sidebar across tools and devices
  • Move sessions between local and cloud without interrupting execution
  • Review code changes using an integrated diff and pull request workflow
  • Use built-in browser and plugins to extend agent capabilities

To use it, update Cursor and open the Agents Window from the command palette.

Try cursor 3
Signals
1 Andrej Karpathy shows LLMs building personal knowledge bases that organize, query, and expand research automatically 25,232 Likes
2 Google launches Agent Skills allowing Gemma 4 to perform tasks across tools and data sources on-device 1,401 Likes
3 Perplexity rolls out Computer in Slack to orchestrate agents, tools, and workflows from one shared interface 1,219 Likes
4 OpenAI enables flexible Codex adoption with usage-based pricing and no rate limits 1,203 Likes
5 Sakana AI opens beta for Marlin, an autonomous assistant that runs hours-long business research tasks independently 993 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