😺Open AI that's ACTUALLY open

PLUS: A weird ChatGPT cover up + the ultimate writing prompt...

Welcome, humans.

One of the most important investigative discoveries of our time has just been uncovered:

For whatever reason, ChatGPT refuses to say the name “David Mayer.” Like, it straight up won't do it. Try it yourself—the chat immediately crashes with a red error message.

Internet sleuths have tried everything to crack this mystery—ciphers, riddles, even Hebrew translations. Nothing works.

Weirdly, other AI chatbots like Gemini and Claude have no problem saying the name. It's just ChatGPT that treats Mr. Mayer like Voldemort.

Two theories have emerged: Either GPT is protecting David Mayer de Rothschild (heir to one of history's largest fortunes) or it's blocking the alias of a Chechen militant that caused a British historian to get accidentally put on a no-fly list.

At least we finally found something ChatGPT won't talk about. Usually getting it to stop explaining things is the hard part!

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

New paper highlights the problems with today’s “open source” AI.
Musk sought to block OpenAI's for-profit transition in court.
Adobe created a tool that makes custom sound effects.
Philippine office workers lead global AI adoption.

Advertise in The Neuron here.

New research shows the problem with “Open Source” AI is that it’s not that open…

Open source AI isn't as open as you might think—and that's a problem, according to a new paper in Nature.

See, when AI companies like Meta call their models “open source,” they're often just sharing the end result (model weights) while keeping the most important parts closed: the training data, code, and infrastructure needed to actually build these systems.

Here's what that looks like in practice:

Meta's LLaMA-3 was trained on 15 trillion tokens, but nobody outside Meta knows what data was used or how it was processed. That's like giving someone a cake recipe without ingredients or instructions. For our Great British Bake Off fans out there: talk about a tough technical!

But there's an even bigger issue—the research shows that building and deploying AI at scale requires three key resources that are controlled by just a handful of companies:

Computing power (NVIDIA controls up to 90% of AI chips).
Training infrastructure (you need NVIDIA's proprietary CUDA framework).
Cloud platforms (deployment usually requires big tech partnerships).

Consider Mistral AI, the lovable French AI underdog. Despite raising over €1B as an open source AI company, they still had to partner with Microsoft Azure to reach users effectively. Why? Because the “open source” label doesn't solve the underlying infrastructure problem.

“Unless pursued alongside other strong measures to address the concentration of power in AI, including antitrust enforcement and data privacy protections, the pursuit of openness on its own will be unlikely to yield much benefit…”

Warning from the researchers

Enter OLMo 2. Last week, AI2 (Allen Institute for AI) released OLMo 2, demonstrated truly open source A by releasing:

Complete training code and data.
500+ training checkpoints.
Full evaluation suites.
…and everything under the Apache 2.0 license.

The 7B model of OLMo outperforms Meta's Llama 3.1 8B on academic tests, while the 13B version matches Qwen 2.5 7B using less computing power.

And OLMO is just one of many models AI2 has open-sourced. AI2 also released Tulu 3 for post-training (which helps make AI models better at specific tasks and safer to use). Combined with OLMO, this creates a complete, independent pipeline for building and customizing AI models. You can chat directly with OLMo or Tulu here.

This matters because it gives organizations—especially those handling sensitive data like healthcare companies—a way to build AI models without needing big tech at all.

Our take: Some, like Nobel Laureate Geoffrey Hinton, compare open-source AI to “letting people buy nuclear weapons at Radio Shack.” But efforts like AI2’s mean we're finally seeing what a fully open alternative to big tech's AI stack could look like.

FROM OUR PARTNERS

Want more of your time back? This tool saved sales reps 10 hours per week on CRM data entry.

From Attention’s Aspire case study here

Tired of your sales team wasting time on manual data entry? Aspire's global sales team increased close rates by 70% after implementing Attention's AI-powered sales platform.

The secret?

Automatic CRM updates.
Multilingual call translations.
Performance scorecards that actually work.

Now, Aspire’s sales forecast accuracy is up to 85%, and reps save 10+ hours weekly on busywork.

Check out Attention and see why Aspire's VP of Sales calls Attention a “total game-changer” for their 850-person team.

Treats To Try.

AI Agents List helps you discover and compare agents across categories like coding, research, and productivity.
Foundry helps you build and improve AI agents with real-time human feedback to ensure reliable, high-quality outputs.
Infinite Canvas by Mindpal runs multiple AI agents side-by-side, letting you branch conversations and chain their outputs together (video explainer).
Crono finds qualified B2B leads and close deals faster by bringing all your sales tools and CRM data into one intuitive platform.
EasyChef turns ingredients you have at home into instant recipe suggestions, helping you reduce food waste and cook healthy meals (iOS, Play Store).
Craft.do organizes your notes, tasks and events into nested documents that sync across all your devices.
Muku.ai turns your product URLs into video ads by instantly generating UGC-style content.
Kroto creates and translates video tutorials and documentation into 60+ languages while preserving the original voice and quality.
RedactAI creates personalized LinkedIn posts that match your writing style from any content (tools like this are why 50+% of LinkedIn posts are AI btw…so if you can’t beat ‘em, join ‘em??).
Neuron (no relation) provides a private computer for chat and image generation at home, while earning you passive income by providing GPU power to networks when idle (probs overpriced, FYI…had to include cause the name!)

See our top 51 AI Tools for Business here!

Around the Horn.

Video is a demo of Kling 1.5 Motion Brush… looks better than it works in practice atm, but it shows potential.

Elon Musk filed in court to stop OpenAI's transition to a for-profit company, claiming irreparable harm and anticompetitive behavior.
Adobe researchers developed MultiFoley, an AI model that generates high-quality sound effects for videos using text, audio, and video inputs (cool demos here).
Amazon built a new genAI model called Olympus that can process images and videos to search video archives (could announce it at AWS re:Invent this week).
Indigenous engineers launched AI initiatives to preserve 200+ endangered Native American languages, with programs like First Languages AI Reality leading preservation efforts.
86% of white-collar workers in the Philippines now use AI, as BPO workers report increased monitoring and workloads while facing potential displacement.

FROM OUR PARTNERS

Want to build AI projects faster and cheaper?

Want to build AI projects faster and cheaper? Join Speed Read AI's founders and Dell's experts on Dec 12th to discover how to run AI locally with Precision workstations and NVIDIA RTX™.