Welcome, humans.

So, apparently Gemini is ALSO missing the O.G. GPT-4o model, because it’s apparently calling itself “a disgrace” and getting stuck in depressive self-destructive loops (yo, are we SURE these AI aren’t sentient? Sounds pretty human coded to me!!).

Google DeepMind’s Logan Kilpatrick calls this an “annoying infinite looping bug” and Google vows to fix it. Reminds us of when you’re vibe coding and ask an AI to fix something, it says it fixes it, and then…seemingly doesn’t do anything, over and over?? But, y’know, now it also belittles itself for not being able to pull up from the spiral. Now that I think about it, that does sound a lot like human depression!

Here’s what happened in AI today:

• We compare and contrast Gemini and Claude’s new memory features.

• Anthropic hired Humanloop team.

• DeepSeek R2 is rumored to debut by end of August.

• Beijing opened the worlds first humanoid “Robot Mall.”

Three Story Thursday…

As you probably know by now from reading The Neuron, the AI industry moves FAST. So fast, in fact, it’s nearly impossible to keep up with it all (*trust us, we try).

We kept opening our laptops thinking we'd write about one big story, but instead kept falling down rabbit holes that led to completely different places. By the end of the day, we had three separate deep dives sitting in our drafts folder. Each one tackles a different side of AI that's... well, kinda wild when you think about it.

So instead of picking one, we wrote three. Today, you choose your own adventure:

Option 1: Everyone's rushing to clone Google's Genie 3… Google just dropped a tool that lets you play inside AI-generated worlds in real-time. Think video games, except the entire level gets created from a text prompt while you're playing. Naturally, everyone from Tencent to random open-source developers are now racing to build their own versions (and that open source one is sick). We break down these playable world models and a bit about how they work in this deep dive.

Option 2: AI is making doctors worse at their jobs… A new study in The Lancet found that experienced doctors who used AI to spot cancer actually got worse at finding it when the AI wasn't around. Meanwhile, another paper claims GPT-5 just achieved “superhuman” medical performance. So which is it—are we training better doctors or creating a generation that can't function without their digital sidekick?

Option 3: Why AI code feels “almost right” but never quite works… New research from METR finally explains why developers feel 20% faster with AI but are actually 19% slower. Turns out AI can write code that passes all the tests but still takes humans 26 minutes to fix before it's actually usable. The problem? AI lacks the “big picture” context that makes code production-ready… but the AI Labs are working on it!

FROM OUR PARTNERS

How Canva's Magic Studio Makes AI Work for You

Picture this: Your boss drops a last-minute presentation on your desk. International audience. Needs translation. Oh, and make it look “professional but approachable”—whatever that means.

Most AI tools would have you writing prompts like you're coding a spaceship. 

Magic Studio? Just click and watch it happen: 

• Need copy? Magic Write gets it. 

• Translation? One click covers 100+ languages. 

• Redesign those boring slides? Magic Switch transforms them instantly.

The new Canva Sheets even makes sense of your messy data without requiring a statistics PhD. No prompt engineering. No YouTube tutorials. Just AI that actually works the way your brain does.

Explore Magic Studio

Prompt Tip of the Day.

Tired of your 832,000-row CSV crashing Google Sheets and hitting AI token limits? Here's a clever workaround our very own Web Webster discovered from needing to analyze his own massive dataset.

The trick: Zip your CSV file and feed the compressed version to ChatGPT or Claude. Then tell the AI to “ingest this and break it into smaller files with manageable row counts that you can still work with effectively.”

This technique works because compressed CSV files can be 70-90% smaller than the originals, and AI tools can programmatically extract and split them using built-in libraries. Once you get your smaller chunks, ask the same analytical questions to each file separately, then have the AI combine all the results.

This bypasses the context window limitations that normally force you to truncate datasets, letting you analyze your complete dataset without losing any data.

Instead of fighting file size limits, you're essentially turning the AI into your personal data preprocessing assistant that handles the heavy lifting before analysis.

Treats to Try.

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

1. *Guidde converts your screen recordings into professional video tutorials with AI-generated step-by-step narration and voiceover in 100+ languages.

2. Profound gets your brand mentioned more in AI search results and shows you exactly what ChatGPT says about your company (raised $35M).

3. Gemini is more personal now by learning from your past chats (on by default) and launched Temporary Chats that auto‑delete after 72 hours (for select countries).

4. Conductor runs multiple Claude Code agents in parallel on your codebase, so you can have one Claude fixing bugs while another writes tests and a third adds new features, all working simultaneously in isolated workspaces.

5. Pally tracks all your contacts across social platforms and helps you search through your network.

6. MCP Use turns building custom AI agents into one line of code instead of managing servers and configs yourself.

7. Autumn simplifies Stripe billing into 3 simple API calls for subscriptions and usage tracking—free to try.

8. You can now watch GPT-5 play Pokemon Red… so far, and GPT-5 has also made 2x as much as GPT o3 on Vending-Bench, a benchmark that tracks how well AI models can run vending machine businesses.

Around the Horn.

• Apple’s AI plan hinges on a tabletop robot (J595, which resembles the Pixar lamp) which it’ll release in 2027, a new lifelike Siri (named Bubbles), and home security cameras (J450) that’ll operate your smart house via facial recognition and sensors (so like turning off lights when someone leaves a room).

• DeepSeek R2 is rumored to launch between August 15-30, powered by Huawei's Ascend 910B chips with 82% utilization delivering 91% of NVIDIA A100 cluster performance at 97% lower cost, using a hybrid Mixture of Experts architecture with 1.2T total parameters and 78B active per token.

• Perplexity offered to buy The Browser Co. (who make Dia), Brave, and yes, even Google Chrome (though TBH Perplexity would probably need to get acquired by Apple before it could afford to buy at the rate it offered).

• Waymo now has a new integration with Spotify so you can stream your own music when you’re having a robotaxi drive you around.

• Anthropic acqui-hired Humanloop's three co-founders and about a dozen engineers to strengthen its enterprise AI strategy and compete with OpenAI and Google DeepMind in tooling capabilities.

• A new “Robot Mall” opened in Beijing that’s billed as the first 4S‑style humanoid robot store with 40+ brands, coverage showed prices from ~$278 to ~$278K.

• Normally we save think pieces for Intelligent Insights, but this piece from SemiAnalysis breaks down how GPT-5’s new router is actually the missing link to help monetize ChatGPT’s users with “agentic purchasing”… more or less cutting search out of the equation entirely when shopping online.

FROM OUR PARTNERS

Expert guide on AI evaluations

How are you evaluating your AI outputs? Learn how the experts quickly and accurately evaluate AI using LLM judges! Enjoy 70 pages of content on:

• How to automate evaluations to score, explain, and flag quality issues.

• Advanced techniques like token-level scoring, Chain-of-Thought, and pairwise comparison.

• Practical frameworks and code examples for building your own LLM judges.

Get the free eBook

Thursday Trivia

One is real, and one is AI. Which is which? (vote below!)

A

B

How well did you do on last month’s trivia? Check out ALL the answers and results here… you might even see your own comment featured!

A Cat’s Commentary.

Trivia answer: A is real, and B is AI (P.S: lots of good tips for how to spot AI in B’s thread)!