😺 GPT-4.5 is incoming, but GPT-4o just went WILD.

Welcome, humans.

Y’all remember when Snapchat filters hit the scene, and fundamentally changed the selfie? Well, we’re seeing some early signs that “AI filters” could soon do the same.

Yesterday, we briefly mentioned Pika, one of the top AI video generators, then lo and behold, they just launched something new: It’s called Pikaswaps, and it’s basically the genAI equivalent of the Snapchat filter. Check it out:

Apparently, Pika already has 12M+ users (WILD!), and just dropped its first iOS app to create mind-bending videos on the go. Think turning selfies into memes, adding fantastical elements (like Yoda) to real footage, or transforming photos into dynamic videos with just a text prompt.

Oh, and btw, don’t forget that Google also just made it so creators can use its Veo 2 video generator to create background filters for videos. Friendly reminder to no longer trust anything you see!

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

• GPT-4.5 is imminent and ChatGPT now has 400M monthly active users.
• Figure demo’d robots working together using a single AI system.
• Spotify added ElevenLabs AI voices for audiobooks.
• Sakana AI made GPU code run up to 100x faster.

Advertise in The Neuron here.

GPT-4o is… different, but at least GPT-4.5 is coming any day now…

Only looking at the headlines, you could argue OpenAI hasn’t had the best two weeks; first, Elon tried to disrupt its non-profit to for-profit transition by offering to purchase the non-profit. Then, Elon released Grok 3 and claimed it achieved early benchmarks outperforming OpenAI’s models.

Even OpenAI’s new benchmark for freelance software engineers was one-upped by Claude Sonnet.

While all this might* have caused some scrambling behind the scenes, it doesn’t seem to have impacted the core business of selling AI.

Because by all accounts, OpenAI is crushing it: As OpenAI’s COO Brad Lightcap says, ChatGPT is now up to 400M weekly active users. And that’s not all:

1. 2M+ business users now wield ChatGPT for work.
2. Reasoning Model API usage increased 5x since o3-mini went live.
3. GPT-5, when it goes live “soon,” will be totally free with no usage caps.

Confused about why OpenAI would give away GPT-5 for free? Turns out Pro and Plus users will get more inference (the time ChatGPT takes to generate its answers) for more complicated tasks.

Right now, OpenAI does this via “low,” “medium,” and “high” settings, but a slider scale like Anthropic plans to release with its new Claude model is a lot more user-friendly.

Speaking of GPT-5—According to a report by The Verge, it looks like Microsoft is prepping OpenAI’s GPT-4.5 and GPT-5 models to go live on its Azure servers as you read this.

GPT-4.5 may be on its way, but a new GPT-4o is already here. As Ethan Mollick shared, something's different about GPT-4o, but OpenAI isn't saying what. Sam hinted at this with a cryptic tweet, but there’s no real evidence in the usual places (LM Leaderboard, OpenAI’s own records).

Here’s what we know about the update:

• GPT-4o seems generally “smarter” and less likely to refuse requests.
• Many specialized custom GPTs (like AI illustrators) are completely broken (Ethan’s Innovation GPT w/ 10K+ users stopped working entirely).
• Text-based GPTs appear mostly unaffected.
• No official documentation or warning was provided about the changes.

This creates a thorny problem for businesses building on OpenAI's tech. Some devs are already jumping ship to open-source alternatives like DeepSeek or Llama, where they can host their own models and avoid surprise updates breaking their products.

Our take: This is the double-edged sword of working with OpenAI. We want better models, but those ~2M business users need stability. With GPT-4.5 and GPT-5 on the horizon, OpenAI will need to figure out how to roll out some pretty big changes without disrupting the ecosystem they've built.

Worth noting: OpenAI released a major update to the Model Spec on February 12th that could have something to do with this…

FROM OUR PARTNERS

💼 Want to become an AI Consultant?

This stat is WILD: The AI consulting market is projected to explode from $6.9 billion to $54.7 billion by 2032. Want to capture this opportunity?

So how do you, an AI enthusiast, become a handsomely compensated AI consultant?

Well, The AI Consultancy Project from Innovating with AI has already helped 300+ students build successful consulting businesses.

Current students love:

• Tools and frameworks to find and serve clients.
• 6-month roadmap to building a 6-figure business.
• Methods to land your first client in as little as 3 days.

As a Neuron reader, you get early access to the next enrollment cycle!

Take advantage of the AI boom—request early access to The AI Consultancy Project ASAP!

Prompt Tip of the Day

In a recent competition hosted by Google, Swedish Prompt Championship finalist Joakim Jardenberg crushed 300+ competitors with a radical approach to prompting AI: treat AI like an eager junior colleague across the table, not a robot.

His winning framework turns conventional prompting wisdom on its head:

• Think of AI as an eager junior colleague sitting across from you, not a digital service.
• Guide it like an intern—stay engaged, answer questions, and give real-time feedback instead of just dropping instructions.
• Match the AI's personality and capabilities like you would a human teammate's strengths.
• Don't over-structure every prompt—sometimes, vague instructions unlock surprising creative solutions.
• Actually talk to your AI (yes, literally dictate prompts via Advanced Voice Mode) instead of typing for more natural communication.

Our favorite insight: Jardenberg discovered he wasn't getting help from AI—he was helping the AI. This mindset shift transformed his results from merely good to championship-caliber. Try it out!

Treats To Try.

1. Gradio, which powers most HuggingFace Spaces, now has Gradio Sketch, a “no-code mode” to help you build + share machine learning apps—just install Gradio and type “gradio sketch” in your terminal (more here).
2. PaliGemma 2 Mix is a new model from Google that looks at your photos and tells you exactly what's in them, from reading store signs to finding specific objects—try it here.
3. Together AI lets you run AI models “4x faster and 11x cheaper” than competitors using custom-built GPU clusters (raised $305M).
4. Mercor sends your single application to thousands of remote jobs after a 20-minute interview (raised $100M).
5. Upsolve adds custom data dashboards to your product so each customer can track exactly what they need (~29 day free trial).
6. Deckd turns your Figma slides into perfectly-formatted presentations.
7. Fleet AI Copilot handles your IT support tasks by automatically solving tech problems and managing equipment needs.
8. Landing helps you create branded landing pages in minutes.
9. Riley gives you personalized parenting advice and helps you track your baby's daily activities with tools built by two project managers (free trial available).

See our top 51 AI Tools for Business here!

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

Around the Horn.

Introducing Helix

• Figure demonstrated its “Helix AI”, a single neural network for all tasks and robots, in a new video that shows how the robots work together to put away groceries.
• Spotify will now accept AI-narrated audiobooks made with ElevenLabs' voices—two years ago, it made a similar deal with Google Play Books.
• Sakana AI created an AI system called The AI CUDA Engineer that makes GPU code run 10-100x faster.

Intelligent Insights

Majorana 1 Explained: The Path to a Million Qubits

• A paper from DeepSeek presented a new attention mechanism that achieved an 11.6x speedup on long text (64K context) by carefully aligning an algorithm with how GPUs actually process data.
• The EU AI Act contains a copyright exemption for “text and data mining” that unexpectedly allows big tech companies to use creative works to train AI without permission or payment.
• A new AI algorithm called Torque clustering can find hidden patterns in complex data without human guidance (read more).
• The EFF argues that expanding copyright to defend against AI is a “snake oil solution” that would cause more problems than it solves, while suggesting focused regulation in specific problem areas (privacy, labor, environment, antitrust) would be more effective to address legit concerns about AI's impact.

A Cat's Commentary.