😺 Tea on the next GPT

PLUS: Teacher dish on how to catch AI cheaters...

Welcome, humans.

Teachers are spilling all the tea on catching students using AI: apparently, every time you ask ChatGPT-4 to “Tell me a story”, it nearly always tells you about a girl named Elara who lives in the woods (and often repeats the same jokes, too).

To test their students, one teacher gave them an open writing assignment—and the instructions specifically stated that the story could NOT be about a girl named Elara, or you’d get -99 points (out of 100)!

Even with those explicit instructions, two students failed. When the students asked about their grade, the teacher explained—BUSTED!

This story inspired more teachers to share other great tactics to catch AI in action, and to improve teaching—there’s even some tricks to help improve your resume’s chances of getting noticed by AI job recruiters!

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

OpenAI’s new ChatGPT model is worrying insiders.
There’s a new version of the music AI Suno.
Someone made a mini AI cluster with new Mac Minis.
Taiwan's chip production soared despite cutting China sales.

Advertise in The Neuron here.

It’s looking like the next version of ChatGPT might not be so good after all…

If you’ve been wondering why OpenAI has announced a slew of cool new features (but not GPT-5) and lost a bunch of key execs (and still not released GPT-5), this might be the reason:

According to a new report, the new GPT code-named Orion isn't showing the same impressive leap in performance we saw from GPT-3 to GPT-4.

Based on first-hand accounts from OpenAI employees who tested it, Orion is better at language but not “reliably better” at other tasks like coding.

Here’s why this is a problem:

Early tests with only 20% of training showed Orion matching GPT-4 quality.
At 100% training, Orion marginally outperforms its predecessor.
This challenges a critical industry assumption called “scaling laws”, or the idea that more compute + more data = consistent exponential AI improvement.

Guess what? OpenAI is running out of BOTH data + compute…

Sam Altman admitted OpenAI faces “hard decisions” about compute allocation.
Employees say they've “largely squeezed as much out” of available internet data.
The data centers needed could cost $100B each, though OpenAI argues a 5GW center could contribute up to $20B in GDP.

Even VCs at A16z are saying current AI models are “not getting the intelligence improvements at all” with additional compute:

New AI Policy Update on Safety, Censorship & Unexpected Risks

Hearing this news, longtime AI critics took a victory lap. However, Amir Efrati suggests the report (which he wrote) actually points to a new scaling paradigm using reasoning models like o1.

Example: OpenAI researcher Noam Brown says the current scaling laws will stop making financial sense at a certain point:

“…are we really going to train models that cost hundreds of billions of dollars, or trillion of dollars?”

Noam Brown, AI researcher talking at the TED AI conference.

So now OpenAI has a new team working on improving models without new data, attempting to expand on o1's success. Brown says having a bot think for just 20 seconds in a poker game matched the performance of “scaling up the model by 100,000x and training it for 100,000 times longer.”

Oh, and we’re “just at the very beginning of scaling up in this direction.”

FROM OUR PARTNERS

Invest in RAD Intel: Limited Opportunity at Prime Valuation

Here's the thing about marketing in 2024: Companies spend billions trying to guess what content will resonate with customers. Most of it doesn't work.

RAD Intel flips this model on its head. Their AI analyzes what actually drives engagement, then helps brands like Hasbro and Sweetgreen create content that converts.

The results? Up to 3.5X ROI improvement and marketing costs cut nearly in half.

The numbers tell the story: Their tech is working so well that their valuation has jumped from $5M to $85M in just three years*. In fact, they've already raised $35M from Fidelity Ventures and executives at Google, Meta, and Amazon.

Now, investments are open to everyone, with shares at $0.50. But here's the catch—the raise closes November 21st.

Don’t miss out—invest today!

This is a paid advertisement for RAD Intel’s Regulation CF offering. Please read the offering circular and related risks at invest.radintel.ai. This valuation has been set by RAD Intel. There is currently no public market for the Company’s Common Stock.

Treats To Try.

Ritchie School Alumni: Brittany, Don, & Gabor Share Their Data Science Program Experience

*Want to launch your AI career while skiing the Rockies? University of Denver's MS in Data Science program teaches the hottest skills in tech: deep learning, neural networks, and machine learning. Available on campus or online. Skiing + AI?! Sign us up!
Exo connects your devices (iPhone, iPad, laptop, etc) to run AI models together when you don't have access to high-end graphics cards (see below for an example in action).
Invofox verifies and processes data from any document type - like invoices or bank statements - and turns them into structured, accurate data.
SellRaze lets you create and manage marketplace listings for your items in seconds across eBay, Facebook, and more.
UnifyApps builds custom AI agents that connect to your enterprise data and automate workflows across 350+ business apps (raised $20M).
PaperGen helps you write long-form content (case studies, research papers, etc) with auto-generated charts, citations, and references.
Latta records user sessions and automatically fixes software bugs without requiring you to look at code.

See our top 51 AI Tools for Business here!

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

FROM OUR PARTNERS

The fastest way to build AI apps

Writer is the full-stack generative AI platform for enterprises. Quickly and easily build and deploy AI apps with Writer AI Studio, a suite of developer tools fully integrated with our LLMs, graph-based RAG, AI guardrails, and more.

Use Writer Framework to build Python AI apps with drag-and-drop UI creation, our API and SDKs to integrate AI into your existing codebase, or intuitive no-code tools for business users.

Start building with AI Studio

Around the Horn.

This is probably the best sounding AI music we’ve heard so far… sounds like Suno V4 is going to be next level.

ChatGPT apparently told ~2M ppl who asked for election news to get it somewhere else, while Perplexity got ~4M views for its Election Hub resource.
Chegg, the 24/7 homework helper, lost 500K subscribers and 99% of its 2021 stock value since ChatGPT’s launch.
AI demand has driven Taiwan’s chip production to a record $165B in value in 2024—and that’s despite Taiwan’s biggest chip producer saying it will stop supplying advanced chips to all of its Chinese customers on Monday.

Sunday Special

Check out this guy who made his own mini AI cluster with 4x of Apple’s new M4 Mac Minis:

The setup, barely larger than a few iPhones, can handle open-source AI like NVIDIA's Nemotron 70B and even Meta's massive Llama 405B. At around $8K (assuming the Minis are 64GB), it's not cheap, but it could offer an interesting middle ground between cloud services and traditional GPU clusters—full benchmarks coming soon.