😺Gemma is all you need

Welcome, humans.

DUN DUN! Law & Order fans, rejoice! Dick Wolf's son just launched a new startup called Wolf Games to let true crime junkies solve their own AI-generated murder mysteries. Their first game “Public Eye” drops this summer (trailer).

We’ve already got our first pitch: “Death by Deadlines: A True Crime Story About Someone Who Tried to Read Every One of Their 147 Open Tabs (And Failed).”

Here’s what you need to know about AI today:

• Google released Gemma 3, its best new model yet.
• Anthropic's CEO predicted 90% AI-generated code in 3-6 months.
• USPTO withdrew its AI strategy after Trump administration policy shifts.
• Alibaba released R1-Omni, an emotion recognition model.

Google released a new version of Gemma, its small and open model family.

Google has been quiet on the AI model release front for the last few weeks (which = ages in AI years), and now we know why…

It’s been cooking up Gemma 3, what it calls “the most capable model you can run on a single GPU or TPU.”

The release is actually a family of open models you can run on your own machine, and the new line-up includes models in four sizes (1B, 4B, 12B, and 27B).

What makes Gemma 3 special? It's built on the same research powering Google's Gemini 2.0 models, but designed to run locally on consumer hardware—from phones and laptops to gaming PCs.

The 1B model runs on practically anything, 4B fits on smartphones, 12B works on midrange NVIDIA chips like a 3060, and the 27B runs on higher-end consumer cards.

Here’s why developers like it, according to Philip Kiely, lead developer advocate at Baseten (who is already hosting Gemma 3):

In fact, Gemma 3 actually set a new benchmark for an open model of its size.

Gemma 3 can see and understand images. This combo of vision, brainpower, and smaller size means you can now do locally what used to require expensive cloud services. It also aligns perfectly with Google's push for on-device AI.

Since Gemma 3 fits on smartphones, we could soon see offline visual search, instant camera translations, and AI that "sees" what you're looking at—all while keeping your data on your device.

You can use Gemma right now in Google AI Studio in your browser, download the models yourself here, or run them on the cloud.

Beyond phones, there will be bots. In classic “one more thing” fashion, Google also released Gemini Robotics, a system based on Gemini 2.0 that controls physical robots.

This system comes in two forms: a model for directly controlling robots, and another with advanced spatial understanding for roboticists. Also, it handles three key challenges that have historically made robotics difficult:

1. Generality: It adapts to new situations and solves tasks it's never seen before.
2. Interactivity: It understands conversational commands and adjusts on the fly when things change.
3. Dexterity: It performs complex tasks requiring precise manipulation—from origami folding to packing snacks.

Google's partnering with Apptronik on humanoid robots (video), plus working with Boston Dynamics and others as trusted testers. Here’s a demo of Google’s team casually chatting with a timing belt—because why not?

FROM OUR PARTNERS

Here’s one AI tool we use every week at The Neuron.

A little well-kept secret for companies crushing it with AI is that only a handful of AI tools are actually worth using.

That’s why we—Noah & Grant—use ChatGPT and Claude, along with a killer product called Attention.

Here’s how it works:

1. We tell Attention exactly what to pay attention to in our meeting (goals, budget, etc).
2. Attention listens in to our call.
3. After, Attention outputs important insights, action items, and follow-up emails.

There’s plenty more features, too. So if you sell anything, literally drop what you’re doing and go book a demo with Attention ASAP.

Prompt Tip of the Day

This thread has five great tips for how to get the most out of your prompts.

The TL;DR:

1. Use frameworks instead of vague requests (give AI fill-in-the-blank template outlines—we do this to help us write!).
2. Try the “Lazy Essay” trick (4-part prompt: Assignment, Quotes, Notes, Instructions).
3. Never accept the first answer (refine responses with follow-up prompts).
4. Force AI to pick a side (it's too neutral otherwise).
5. Fix bad responses by changing one variable at a time.

Treats To Try.

1. *Flow for Windows is finally here! Speak 3x faster than typing with AI-powered dictation that works everywhere. Try it today.
2. AI Renamer instantly renames files in bulk to keep your folders organized without the manual hassle.
3. Guse is a spreadsheet that automates analysis, formulas, and insights.
4. Unsloth helps you fine-tune and train language models faster and cheaper—here’s Unsloth’s advice for how to run Gemma 3.
5. ColorMatch harmonizes colors across images for designers who need seamless palettes in a click.
6. Same.dev is a free coding agent that’s really good at cloning website UIs.
7. Four for the enterprise:
   1. Credo AI helps you build trust in your AI systems by streamlining governance, risk management, and compliance across your organization.
      1. The team just released its AI Model Trust Scores, which compares capability, safety, affordability, and speed across 90+ scenarios for your specific industry use cases.
   2. LILT is like a Babel Fish for the enterprise—it unites AI and human expertise to translate massive amounts of content across 100+ languages, and integrates w/ 60+ business tools.
   3. AI21 Maestro breaks down and executes your complex tasks by analyzing strategies, creating tailored plans, and validating results against your requirements.
   4. Moveworks instantly finds information and automates workflows across your entire company (acquired by ServiceNow for $2.9B).

See our top 51 AI Tools for Business here!

*This is sponsored content. Advertise in The Neuron here.

Around the Horn.

• Anthropic CEO Dario Amodei said he believes we have 3-6 months before 90% of code is made by AI, and that spies may be trying to get their hands on his $100M “algorithmic secrets” (official doc).
• The USPTO has withdrawn its AI strategy document, originally published in January 2025, following a shift in federal AI policy under the new administration.
• Smartphone makers are racing to make AI your “second brain,” integrating advanced features to handle tasks like scheduling and personalized recommendations.
• Browser Use got a huge traffic boost from the success of Manus, with daily downloads rising nearly 4x in a week.
• Alibaba released R1-Omni, an emotion recognition model using reinforcement learning that showed superior reasoning and generalization abilities.

FROM OUR PARTNERS

Google Cloud’s Future of AI: Perspectives for Startups

Want unique perspectives on the most critical cloud and AI trends, opportunities, investment areas, and challenges that startup leaders need to know in 2025??

Hear from 23 industry leaders, including Amin Vahdat (Google Cloud), David Friedberg (Ohalo Genetics), Harrison Chase (LangChain), Crystal Huang (GV), and more.

Download here

Thursday Trivia

One is real, and one is AI. Which is which? (vote below!)

A.

B.

Which is AI?

The answer is below, but place your vote to see how your guess compares to everyone else (no cheating now!)

Here are the results from last week’s trivia:

Here’s what you said:

• D.R chose B: The blurry strawberry on the table doesn't look natural
• D.L chose A: It’s getting hard and hard harder to tell
• M.D chose B: I always guess whichever one looks better to be AI

A Cat's Commentary.