😺 Here's OpenAI's plan to fix AI hallucinations...

PLUS: Taco Bell's AI drive thru goes waaaay wrong...

Welcome, humans.

Remember our live stream on LM Studio from last week? We turned it into a blog! Learn everything there is to know about how to get started running AI models on your own computer for free (and the cool tools that help you do this), along with helpful time-codes to track the key parts of the convo if you want to skim through the episode itself.

P.S: Gonna do another one of these on Thursday! This week’s topic is: Text-to-Speech Throwdown: Testing Hugging Face’s TTS Arena V2 on The Neuron LIVE! Just click the “remind me” button on YT to get a reminder notification when we actually go live.

Now, a story just went viral on Reddit about how a tiny Caribbean island accidentally became one of the biggest winners of the AI boom. Check it out:

Back in the 1980s, Anguilla (population: oh, ~16,000) got assigned the .ai domain when countries were getting their two-letter codes.

Nobody thought much of it at the time (the internet was barely a thing, and “AI” obviously wasn't on anyone's radar. But now? Every AI startup wants a .ai domain, and they're paying crazy money for it. User Think_Bunch3020 says last year alone, those domain registrations brought Anguilla $39M, nearly a quarter of their national budget!

So, it's basically free money falling from the sky because of a bureaucratic decision made 40 years ago. Obviously, the thread is full of people joking about starting their own countries in anticipation of new tech trends: we need countries for quantum computing (.qc), brain chips (.bc), humanoid robots (.hr?? lol), nuclear fusion (.nf), gene therapy (.gt), augmented reality (.ar), and synthetic biology (.sb).

How about Genetheratopia? Synthoslovakia? Quantum Computistan? The Principality of Brain Chipenstein? The People's Republic of Nuclear Fusionberg?! Now that’s a banger!

Here’s what happened in AI today:

We break down OpenAI’s new hallucination paper.
Supabase (database tool powering Lovable) on track to raise $5 billion.
ElevenLabs approved $6.6B tender offer for staff share sales up to $100M.
Google made Veo 3 video models production-ready with 50% price cuts.

OpenAI Finally Figured Out Why ChatGPT Makes Stuff Up (And How to Fix It)

DEEP DIVE: OpenAI knows how to solve hallucinations… but what about creativity?

Ever asked ChatGPT for a specific fact and gotten three different answers on three different tries? Part of that’s intentional (Probability of next token prediction! Surprise! Yay!) and part of that’s unintentional (hallucinations producing fake information! Boo!).

Well, OpenAI just shared a new research paper explaining why AI models confidently make things up—and more importantly, how we might finally fix it.

First, the problem isn't what you think. It's not about making models smarter or training them on more data. According to OpenAI's new research, hallucinations persist because we're essentially training AI to be terrible test-takers who guess instead of admitting uncertainty.

Here's the deal: When researchers test language models, they use benchmarks that work like multiple-choice exams. Get it right? One point. Say “I don't know?” Zero points. Result? Guess wildly and maybe get lucky? You might score!

Think about it like a student facing a pop quiz. If leaving an answer blank guarantees failure but guessing gives you a 1-in-365 chance of nailing someone's birthday, what would you do? You'd guess. That's exactly what ChatGPT does.

Show me the incentives, and I’ll show you the outcomes, as they say…

OpenAI's research shows this creates a vicious cycle:

Models learn to bluff: Just like students who write something plausible-sounding when stumped, AI models generate confident-sounding nonsense rather than admit uncertainty.
Current benchmarks reward this behavior: Most popular AI evaluations only measure accuracy (how many questions you get right), not whether you're honest about what you don't know.
The problem compounds: Even as models get more capable, they keep hallucinating because that's what scores best on tests.

So what's the fix? OpenAI proposes something surprisingly simple: change how we grade AI tests. Instead of the current all-or-nothing scoring, training needs to penalize wrong answers more than “I don't know” responses AND give partial credit for appropriate uncertainty. Suddenly, admitting you don't know becomes the smarter strategy. Or as Corey calls it, “the humility to shut up.”

Why this matters for you: If implemented across the industry, this could mean AI assistants that actually tell you when they're unsure instead of confidently serving up made-up facts. Less misinformation, more trust, and finally… an AI that knows what it doesn't know. After all, the more you learn, the more you realize how much you don't know…

FROM OUR PARTNERS

Your keyboard called—it’s taking a long weekend ⌨️✈️

We’ve been banging away on keyboards for 150 years. Until now, voice dictation hasn’t been reliable enough to change that.

Wispr Flow finally delivers the no-edit confidence we’ve all been waiting for:

4× quicker than typing. Dictate emails, docs, and DMs in real time and save precious hours every week.
AI auto-edits on the fly. Flow cleans filler words, fixes grammar, and formats perfectly as you speak.
Works inside every app with no setup. Fly through Slack notifications, give more context to ChatGPT, or brain dump into Notion.
Use it at your desk or on the go. Available on Mac, Windows, and iPhone.

“This is the best AI product I’ve used since ChatGPT.”
— Rahul Vohra, CEO, Superhuman

Give your hands a break ➜ start flowing for free today.

Prompt Tip of the Day.

Found a great prompt tip on Reddit yesterday for using AI to help you ask the right question (not just get the right answer):

Using AI + Socratic Method to Find the Right Question

AI is great at giving answers — but its real power is helping you sharpen the question.

Here’s how:

State what you think you know. Feed AI your assumption.

Have it interrogate the “why.” Let the AI keep pressing until the foundation is exposed.

Ask it to invert. “What if the opposite were true?” → see what new possibilities emerge.

Loop this until the noise collapses into one clear, high-value question. That’s the one worth pursuing.

Treats to Try.

*Asterisk = from our partners. Advertise in The Neuron here.

*Chatbase creates support bots that handle customer questions and route tricky issues to humans.
Google's Veo 3 and Veo 3 Fast video models are now stable production-ready through the Gemini API, with prices cut nearly 50% (which means you can include them in your own agent workflows or software); you can also generate 1080p HD videos in vertical 9:16 format (perfect for TikTok and Instagram.
Koah lets you place ads directly inside AI chat conversations when users ask questions related to your business, reaching 500M+ daily GenAI conversations across 50+ apps like Liner (student research), OpenEvidence (search for doctors), and DeepAI (raised $5M).
Higgsfield creates ready-to-share images and videos in one click using AI, with pre-built templates for everything from ASMR content to billboard ads, so you can make polished commercials without any editing skills.
Cognition creates Devin, an autonomous AI software engineer that takes full projects from start to finish, and now owns Windsurf, an AI-powered code editor that assists you while coding makes (raised $400M); also they just hired Swyx so basically they’re legends in our book.
NotebookLM’s Reports feature now allows you to create fully customized reports with personalized structure and style in 80+ languages, provides dynamic topic suggestions based on uploaded sources, enables customizable prompts, and introduces blog posts as a new report template.
Tripo3D turns your text descriptions or photos into ready-to-use 3D models with smart retopology that automatically optimizes them for games, 3D printing, or design projects—free trial, then $29.90/month.

Around the Horn.

Excuse me WHAT (here it is)

Anthropic and the class action lawyers suing it have until Sept 15th to get all the authors covered under the lawsuit to “opt in” to the settlement, or their works won’t be covered by the $1.5B the company plans to pay out.
Supabase, the database tool powering Lovable, is on track to raise $5 billion.
ElevenLabs okayed a $6.6B tender offer, letting staff sell up to $100M of shares.
Amazon launched Weekly Vibe AI playlists that auto‑refreshes every Monday.
In case you missed it, AI now has its own “Darwin awards” for famous fails, and right now, the Taco Bell drive-thru fail that enabled a guy to order 18K waters is leading the nominees.