😺 20-Somethings Break AI 💥

PLUS: More evidence of AI dumbing us down?!

Welcome, humans.

Anthropic's Claude AI just pulled a classic AI move: it completely fabricated a legal citation, and then—plot twist—the lawyers actually used it in court. Awkward.

The scene: A Northern California courtroom, where Anthropic's lawyers are battling music publishers. Claude decides to get creative and generates a citation out of thin air—complete with a made-up title and nonexistent authors. And get this: their “manual citation check” didn't even catch it.

It's like the AI equivalent of writing your homework, but then your dog actually eats it—except in this case, the dog actually wrote the homework, and it doesn't make any sense. Nor does this metaphor, which is increasingly getting away from us…

This is far from the first time AI has fumbled in the legal arena, and it’s not exclusive to Claude. Earlier this week, another judge absolutely roasted law firms for submitting “bogus AI-generated research.” We seemingly see these headlines every week; here’s all the ones Perplexity and OpenAI Deep Research could dig up (remember, these too could have hallucinations in them!).

Anyway, talk about confidence, right? Feels like more of a lawyer problem than an AI problem. Consider this your friendly reminder to always check your work. Twice.

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

Sam Altman exposed how 20 year olds use AI (and it’s good advice).
Google introduced a new coding agent for algorithm discovery.
Salesforce released new image models.
Meta delayed its Behemoth model due to insufficient improvement.

Advertise in The Neuron here.

Sam Altman Unveils How Young People Use AI

OpenAI CEO Sam Altman just dropped some serious insights about how younger generations are using AI—and it's way more sophisticated than your typical Google search.

During a recent talk at Sequoia Capital, Altman revealed that young people aren't just using AI tools; they're basically treating them like a life operating system. Move over, smartphone. There's a new digital assistant in town.

Here's the wild part: While older folks might use ChatGPT as a glorified Google replacement, people in their 20s and 30s are using AI as an actual life advisor. We're talking about a digital confidant that knows the full context of their life.

Altman shared some mind-blowing observations:

Young people are memorizing complex prompts like they're cheat codes.
They're setting up intricate AI systems that connect to multiple files.
Most shockingly? They don't make life decisions without consulting ChatGPT.

OpenAI’s Sam Altman on Building the ‘Core AI Subscription’ for Your Life

That’s right. Young users have developed sophisticated AI workflows:

Connecting multiple data sources.
Creating complex prompt libraries.
Using AI as a contextual advisor that understands their entire social ecosystem.

It's like having a super-intelligent friend who knows everything about your life, can analyze complex situations, and offers personalized advice—all without judgment.

Altman compared this to the smartphone revolution. Remember how younger people instantly got touchscreen tech while older generations struggled? We're seeing the same pattern with AI.

If you want to understand the future of AI, watch how 20-year-olds are using it.

Our take: AI isn't just a tool for these young users—it's becoming an extension of their decision-making process. Whether you find that exciting or slightly terrifying probably depends on your age.

Want to level up your OWN AI game? Start thinking like a 20-year-old: complex prompts, multiple data sources, and AI as your personal advisor. This is a slightly older video (from last September) but it’s a GREAT resource for learning how to use Claude like a 20-y/o.

36 Claude Tips for Beginners in 2024! (Become a PRO!)

Here’s the most useful parts for implementing Sam’s advice:

Creating Projects (47:23): Learn to set up specialized workspaces in Claude for different life areas like the digital natives Sam Altman described.
Custom Instructions (49:00): Master persistent instructions that shape Claude's responses—the "complex prompts" young people memorize.
Project Knowledge (52:04): Upload multiple references (up to 500 pages) to give Claude comprehensive context about your life and decisions.
Combining Features (57:23): Build sophisticated AI systems by connecting instructions, shortcuts, and knowledge bases into a contextual advisor.

By creating projects with custom instructions and connecting multiple data sources, you can transform Claude from a simple search tool into a personal advisor that understands your full context—just like the tech-savvy 20-somethings who wouldn't make the decision to watch the above video or not without consulting their AI first.

FROM OUR PARTNERS

SuperAI Singapore, 18-19 June 2025

Asia’s largest AI and robotics event, SuperAI (www.superai.com) brings together frontier technology visionaries, developers, startups, enterprises, researchers, and policymakers to shape the future.

Taking place 18-19 June 2025 at the iconic Marina Bay Sands, SuperAI Singapore will convene over 7,000 attendees from more than 100 countries to explore and unveil developments in robotics, healthcare, finance – and AI's impact across industries and society.

Use the code SUPERNEURON15 at checkout for 15% off tickets.

SAVE 15% NOW

Prompt Tip of the Day

Can AI Pass a Top-Tier Engineering Coursework?

TBH this is just a fascinating video, and since we threw an hour and a half explainer video at you in the above article, we figured any other prompt tips should be reserved for another day 🙂

This thread summarizes the video. Essentially, an Imperial College engineering professor tested ChatGPT, Claude, Meta AI, and Gemini on complex coursework involving coding and advanced math, with TAs blindly grading the submissions.

Turns out, ChatGPT and Gemini passed the assignment (with Gemini performing significantly better, because duh it’s the best AI atm), while Claude and Meta failed, raising important questions about how technical education must adapt to AI's evolving capabilities. Which is, on net, probably making people dumber.

Treats To Try.

Mistral Small 3.1 is a small AI that can understand images and supports 128K tokens (equivalent to ~96K words or ~200+ pages)—try here or download this base version or instruct version.
Klangio converts your music recordings into editable sheet music as well as digital music formats like TABs, MIDI, and MusicXML files.
Freepik now lets you use Gemini’s image changer (which is going viral for its controversial ability to remove image watermarks) in its AI suite—you can also use it here or in Google AI Studio.
Currents scans Reddit (maybe others? seems mainly Reddit-based) to show you exactly what people are saying about your market, competitors, and trends.
Fluently helps you speak English confidently on work calls by analyzing your speech and creating personalized improvement plans.
Gadget helps you build, run, and scale web applications in hours with AI assistance, built-in integrations, and serverless infrastructure.
Windsurf released SWE-1, its own AI models that supposedly rival Claude 3.5, 4.1, and Gemini 2.5 and speeds up your entire software development by understanding and building on your edits, terminal commands, and browser interactions—free for paid users.

See our top 51 AI Tools for Business here!

Around the Horn.

Google DeepMind introduced AlphaEvolve, a Gemini-powered coding agent for algorithm discovery that's been secretly running for about a year—unlike previous AI systems that use reinforcement learning to optimize model weights, AlphaEvolve modifies, evaluates, and optimizes code directly; more on Monday!
Salesforce released BLIP3-o on Hugging Face, a family of fully open unified multimodal models that generate rich CLIP image features using a diffusion transformer.
LangChain released LangGraph Platform, allowing users to deploy, scale, and manage agents with long-running, stateful workflows.
Meta delayed the release of its new Behemoth model to at least this autumn because apparently it’s not improving well enough or fast enough.
Grok apparently went haywire over the last 24 hours and repeatedly referenced “white genocide in South Africa” even when responding to unrelated topics when @’d on X.com—the company responded, blaming an “unauthorized modification” that violated their “internal policies and core values,” and will now publish all their system prompts under Grok Prompts.

Intelligent Insights

Apparently it’s really easy to make an agent work with just messaging a language model (what ChatGPT is) and giving it tool use (that’s according to Sketch, an open-source AI coding agent).
This article by Vincent Cheng explores how relying on large language models for tasks like coding can erode skills development and learning.
This reasoning model supposedly outperforms DeepSeek-R1 and rivals much larger models like Qwen3 on math and coding benchmarks despite having far fewer parameters (which are like AI brain connections).
Despite early predictions of AI replacing radiologists, the field has flourished with increased employment as AI transforms roles rather than eliminates them—resembling how ATMs changed banking rather than how automobiles replaced horses (side note: isn’t it kinda silly to compare horses losing their jobs to humans?).
A new study shows language models perform 39% worse in multi-turn, underspecified conversations compared to single-turn, fully-specified instructions—mostly due to premature assumptions and overly verbose outputs.