Welcome, humans.

The closer we get to December, the closer we get to more Apple Intelligence features rolling out on a wider scale. This also means it’s time for another friendly reminder that everything in AI is a “work in progress”—to say the least. Example:

One reviewer said Apple Intelligence notification summaries are “honestly pretty bad” because they’re “often wrong, usually odd, sometimes funny, rarely helpful.” Owwwch!  

He categorized the types of summary fails he received from “correct…but totally absurd” (our fave) to “misunderstanding sarcasm and slang”, “loss of context”, “overload”, and “just plain wrong.”

All that said, he decided to leave the summaries on—because they’re “usually inadvertently hilarious.”

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

• Mistral launched a bunch of new open source AI features to try out.

• Biden and Xi agreed to keep AI away from nuclear weapons.

• Swiss chapel replaced priest with AI Jesus.

• The Beatles' “last song” nominated for two Grammys (thanks to AI).

Why everyone's talking about Mistral's new AI…

France's Mistral AI just dropped its biggest release yet: a 124B parameter model that can process images, understand documents, generate artwork, search the web, and even code in real-time. Oh, and they made it all free (on Le Chat) and open-source.

Wait, what?! Ya. FREE.(for now, anyway—try it here while you can). Viva La France!

Mistral’s release is genuinely impressive. Their new Le Chat platform now includes:

• Pixtral Large, a 124B parameter vision model (available via API).

• A Canvas feature (similar to Claude's artifacts) for real-time coding.

• PDF analysis that can understand complex documents and charts.

• Image generation powered by Black Forest Labs' FLUX1.1.

• Web search.

• The ability to create agents for repetitive work (demo here).

This image sums up all these capabilities nicely:

While Le Chat is undoubtedly powerful, there’s a debate over whether the model powering it (Pixtral Large) is really the frontier of open weight models its size. Early benchmarks suggest that Alibaba Cloud's Qwen2-VL 72B—a model almost half the size—might actually outperform it on several key metrics.

This isn't that surprising to those who've been watching Qwen 2.5 closely. The new 2.5 Turbo model (try it here) has been quietly setting records of its own.

And devs in particular like the Qwen 2.5 Coder 32B because it can run surprisingly well on consumer hardware.

This is the point: Ignoring o1, most of the impressive advances in AI lately aren't coming from bigger models, but from more efficient ones.

Take SmolM2 from HuggingFace—a family of tiny AI models, the smallest of which is just 88.2Mb, but can still hold a conversation. The largest version, 1.7B-Instruct, has been called “probably the best under 2B model I have ever tried.”

So what’s going on here? Just last week, researchers from a murderers row of Ivy Leagues published findings that show bigger isn't always better when it comes to AI.

Here’s the TL;DR:

• When you train a large language model, it keeps getting better as you show it more data (duh).

• But if you then try to make that model run faster by lowering its precision (basically, using fewer bits to store numbers), it can actually perform worse than if you'd trained it on less data.

• Think of it like compression: The more information you pack into a model during training, the more sensitive it becomes to any information loss when you try to compress it later.

FROM OUR PARTNERS

Adobe Stock Update

Check this out: New features in Adobe Stock enable creators like you to accelerate your content selection process and find more unique, relevant content you can customize to your aesthetic and brand preferences.

Powered by Adobe Firefly, you can now customize stock images before purchasing: 

• Swap or remove backgrounds with a click. 

• Expand images to fit any format. 

• Generate variations that match your chosen image's style or composition. 

Just type a prompt, and watch as AI transforms your stock picks (while maintaining their professional quality). 

Check it out here.

Treats To Try.

1. *AssemblyAI announced a new milestone with the release of their most advanced speech-to-text model to date. Universal-2 delivers even greater accuracy and precision for impeccable audio data. Try Universal-2 for free here.

2. Quantera helps you research stocks faster by letting you chat with 30 years of financial data and critical details like risk factors.

3. Momen builds fully functional web apps for you, from design to deployment, directly in your browser.

4. SuperAnnotate lets you collect, curate, and analyze training data to fine-tune your language models with specialized annotation teams (raised $53M).

5. Juna.ai controls your factory's production processes to maximize output and reduce energy waste.

6. KitOps assists you package your AI models, datasets, and code into standardized, reproducible bundles that can run anywhere.

7. Carter Pro is a flexible robotic assistant to help warehouse and manufacturing workers move materials w/o needing to build anything new.

8. Layer helps you visualize and plan projects using mind map trees—LOVE this!

See our top 51 AI Tools for Business here!

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

Around the Horn.

• ElevenLabs launched a new Conversational AI feature where you can create your own AI agent to chat with a knowledge-base, that even integrates with your own custom model (video demo here).

• Perplexity previewed a new “shopping” feature for Pro users that creates product cards with pricing, product, and pros and cons on it.

• U.S. President Joe Biden and Chinese President Xi Jinping agreed they won’t let AI take control over nuclear weapons. Oh, good! (…that was on the table?!)

• A chapel in Lucerne, Switzerland created an AI Jesus (trained on the bible, of course) to give advice in the confessional instead of a priest.

• Paul McCartney used AI to fix up and release the Beatles’ “last record” Now and Then, and now the track is up for two Grammy nominations.

FROM OUR PARTNERS

Generative Edits (powered by Adobe Firefly) lets you edit any Adobe Stock image before licensing it: replace or remove backgrounds, expand images, generate variations, apply a new style, or match a composition you love (like above)—check them out here. 

Prompt Tip of the Week

Ever notice how ChatGPT's responses sound a bit... ChatGPT-ish? There's a trick called “echo writing” ppl are using that can help fix that. Instead of getting the usual GPT-flavored phrases, you can make ChatGPT write in your voice.

Here's how to do it:

• Give ChatGPT a sample of your writing and say ‘Write in this style, matching my sentence structure, vocabulary, and tone’

• Create a phrase map: List common ChatGPT phrases (“It's important to note,” “On the other hand”) and your preferred alternatives (“Just remember,” “But flip side”)

• Ask ChatGPT to analyze your writing style first, then use those characteristics in its response (like if you use shorter paragraphs or casual language)

Simple example prompt: “Here's a sample of my writing: [paste your text]. Please write the next response matching this exact style, including my sentence length, vocabulary choices, and overall tone.”

Remember—this works best when you give ChatGPT clear, specific examples of your writing style rather than vague instructions like “write casually.”

Here’s an example one you can copy + edit to your needs.

A Cat's Commentary.