AlphaSignal

Hey ,

Google I/O just handed developers the keys to a fully managed AI stack, from pocket to production. Build an app on your phone, deploy an agent with one API call, and let Google handle the servers. No setup, no sandbox management, no excuses.

The pattern is clear: Google is collapsing the distance between idea and shipped product. The dev stack is getting swallowed whole.

Meanwhile, KAIST's tiny 10M-param GRAM model quietly outperforms giants on hard reasoning. Small and smart is having a moment.

Together with

Today's Author

Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.

Summary

Read time: 6 min 42 sec

Top Repo

▸ GPT-Image-2 hallucinates hidden text in dot patterns that changes every time

ASUS

▸ The Always-On Mini AI Agent

Top News

▸ Google AI Studio is coming to mobile so you can build on the go

Prior Labs

▸ TabPFN-3: Predict anything from structured data

Top News

▸ Google launches managed agents on Gemini API with built-in Linux environments

Signals

1. Meta releases SAM3 open-source, tracking any object in video by text description

2. Presented by Iluvatar Labs: Marvin's open-source science initiative tackles humanity's hardest problems - from schizophrenia to muscular aging.

3. KAIST's 10M-param reasoning model hits 97% on Sudoku and 52% on ARC-AGI-1

4. Stanford study finds large models train better with no data filtering at all

5. Tencent releases open-source translation models that fit in 440 MB on-device

6. NVIDIA ships verified agent skills with built-in security cards for Claude, Codex, and Cursor

Top Repo

GPT-Image-2 hallucinates hidden text in dot patterns that changes every time

4,005 Likes

Someone fed an image of orange dots to GPT-image-2 and the model swore there was hidden text inside the pattern. Except there wasn't. No secret message, no steganography (data hidden inside an image). Just dots.

Here's the weird part: the model is totally confident, but reads a different message every time. It's a visual hallucination, meaning the AI sees patterns that aren't there, like a brain inventing shapes in clouds. And it only triggers via GPT-image-2, not through direct prompting.

Bonus chaos: Gemini models consistently read the same image as "SEND NUDES." Every time. Reproducibly.

Out of this rabbit hole came GraphGlyph, a Python tool that actually encodes real text into these dot-graph images. You can:

Encode any string into a blue-and-orange graph image
Decode it back perfectly from the SVG output
Run it yourself: clone the repo and use graph_cipher.py encode/decode

The takeaway: AI vision models pattern-match so hard they invent meaning from noise. That's both a bug and, apparently, an art form.

forward →

Presented by ASUS

The Always-On Mini AI Agent

The ASUS NUC 16 Pro delivers 180 TOPs of AI performance with Intel Core Ultra and Arc. Built for your everyday agentic AI assistant.

Supports next-gen agentic AI apps like OpenClaw and Hermes Agent
Automate everyday tasks like cross-platform file organization
Designed for 24/7 reliability and consistent performance

Enable a smarter AI assistant experience on the ASUS NUC platform.

GET ASUS NUC

partner with us →

Top News

Google AI Studio is coming to mobile so you can build on the go

3,043 Likes

Google just announced AI Studio is coming to your phone as a standalone app. Think of AI Studio as Google's browser tool where you describe an app idea and it builds it for you. Now that power fits in your pocket.

Here's what you can actually do with it:

Speak or type your idea and watch it generate a working app on your phone
Preview builds directly from mobile, then go deeper when you're back at your desk
Remix apps from a community gallery for inspiration
Share live deployments with friends to collect feedback
It builds in the background, so you don't have to keep the app open

The iOS version has a July 1 release date spotted on the App Store, with Android still showing "coming soon." It supports 76 languages and needs iOS 18 or later. Pre-registration is open now on both stores.

forward →

Presented by Prior Labs

TabPFN-3: Predict anything from structured data

Still building custom pipelines for every prediction task?

TabPFN-3 by Prior Labs does it in one call, no training, no tuning.

93% win rate over 8-hour tuned XGBoost and LightGBM
Churn, fraud, pricing, demand forecasting - any prediction task
0.2s inference on 1M rows

TRY TABPFN NOW

partner with us →

Top News

Google launches managed agents on Gemini API with built-in Linux environments

2,684 Likes

Building an AI agent used to mean setting up your own servers, managing sandboxes, and writing a ton of infrastructure code just to get something running. Google just removed most of that work.

Meet Managed Agents on the Gemini API. With a single call, you can spin up an agent that reasons, uses tools, and executes code in an isolated Linux environment. Google hosts and runs everything.

Here is what the agent can actually do out of the box:

Execute code, run Google searches, and read URLs by default
Manage files and browse the web to retrieve current information
Retain files and state between calls, so you can resume the same session later

Want to customize it? Instead of writing complex orchestration code, you define everything in markdown files like AGENTS.md and SKILL.md, then register them as a managed agent.

On pricing: sandbox computing is free during preview, but token usage for the underlying model and tools is billed at standard rates.

You can try it now via the Interactions API or Google AI Studio.

forward →

Signals

Meta releases SAM3 open-source, tracking any object in video by text description

2,660 Likes

Marvin's open-source science initiative tackles humanity's hardest problems - from schizophrenia to muscular aging.

Presented by Iluvatar Labs

KAIST's 10M-param reasoning model hits 97% on Sudoku and 52% on ARC-AGI-1

1,333 Likes

Stanford study finds large models train better with no data filtering at all

1,029 Likes

Tencent releases open-source translation models that fit in 440 MB on-device

1,823 Downloads

NVIDIA ships verified agent skills with built-in security cards for Claude, Codex, and Cursor

1,626 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

How was today's email?

Awesome Decent Not Great

unsubscribe_me(): return True

Google launches managed agents with Linux environments (4K+ likes)