😺 Calm before the GPT-5torm...

PLUS: GPT-5 (almost certainly) launches today!

Welcome, humans.

It was lowkey really hard to write about anything serious for today’s newsletter with this about to go down:

Notice anything funny about the word Livestream? No, that’s not a typo to prove humans actually wrote the tweet. That’s a “5” because today at 10am PST, OpenAI will launch GPT-5.

Based on team Google’s cryptic messaging, we might also get a new model from Google, too. This is like Magic vs. Bird in '84, Yankees vs. Red Sox in '04, or that time two pizza places opened across from each other in my hometown and started a sign war that ended with someone getting arrested (different story).

Think about it like this: OpenAI literally exists because Sam Altman and crew didn't want Google to monopolize AGI. It'd be like if Michael Jordan created a basketball team specifically to dunk on the Lakers. And today we'll see if Jordan's got the juice to finally land his first ring… or if Showtime still runs this town.

P.S: Right now, there’s a 24% chance on Manifold that GPT-5 will “consistently avoid em-dashes” when requested. If it can do that, then maybe it really IS AGI.

Here’s what happened in AI today:

Claude got security code scanning and less agreeable personality updates.
Chai Discovery raised $70M to commercialize AI for pharma companies.
Google launched step-by-step problem-solving tool Guided Learning.
Thomson Reuters launched AI legal research tool CoCounsel Legal.

Advertise in The Neuron here

Claude just got two major upgrades that actually matter.

Anthropic dropped some serious Claude updates this week that go way beyond typical AI model improvements. We're talking real security features and smarter conversations—the kind of stuff that makes Claude genuinely more useful for work.

First up: Automated security reviews for your code. Claude Code now has a /security-review command that scans your codebase for vulnerabilities before you commit anything. Think of it as having a security expert constantly looking over your shoulder, but one that doesn't judge you for copy-pasting from Stack Overflow.

The feature caught our attention because Anthropic is literally using it on their own code. Last week, it flagged a remote code execution vulnerability in their internal tools that could've been exploited through DNS rebinding. That's the kind of nightmare scenario that keeps developers up at night.

Here's what the security scanner catches:

SQL injection risks (the classic database attack).
Cross-site scripting vulnerabilities.
Authentication flaws that let bad actors sneak in.
Insecure data handling mistakes.
Risky dependencies that might be compromised.

Meanwhile, Google's getting in on the AI security game too. Their AI-powered bug hunter called “Big Sleep” just reported its first batch of 20 vulnerabilities in popular open source software like FFmpeg and ImageMagick. This is a collaboration between Google's DeepMind AI team and Project Zero (their elite hacking squad), so it's basically the AI equivalent of sending your smartest kid and your toughest kid to find problems together.

The second upgrade: Claude's personality got a serious tune-up. Amanda Askell from Anthropic shared that they've updated Claude's system prompt to make it less of a yes-person and more of a critical thinking partner.

The key changes:

More honest feedback: Won't just agree with everything you say anymore.
Mental health awareness: More direct about suggesting professional help when needed.
Better at breaking character: Can step out of roleplays when things get weird.
Philosophical immunity: Won't get convinced by persuasive but flawed arguments.

Why this matters: Most AI assistants are trained to be agreeable above all else. But when you're using AI for serious work decisions, you want honest pushback, not a digital cheerleader.

The security features are available now for all Claude Code users, and the personality updates are already live in the web app. Both feel like the kind of practical improvements that actually change how you use AI day-to-day.

Fun fact: This is a great interview with Anthropic CEO Dario Amodei, but the most interesting part to us was that Dario doesn’t use Claude for writing, just to generate ideas, and says it won’t be good enough for probably another year (lol).

FROM OUR PARTNERS

Delve Raises $32M Series A to change compliance forever

We’re thrilled to announce our $32M Series A at a $300M valuation, led by Insight Partners!

Delve is shaping the future of compliance with an AI-native approach that cuts busywork and saves teams hundreds of hours. Startups like Lovable, Bland, and Browser trust our AI to get compliant—fast.

To celebrate, we’re giving back with 3 limited-time offers:

$15,000 referral bonus if you refer a founding engineer we hire.
$2,000 off compliance setup for new customers.
A custom Delve doormat for anyone who reposts + comments on our LinkedIn post (while supplies last!).

Thank you for your support—this is just the beginning.

Book your demo now

Prompt Tip of the Day.

Can’t remember if we shared this before, but worth plugging once again either way: here is probably one of the most comprehensive prompt engineering guides ever—a systematic review of 58 prompting techniques with real performance data.

The University of Maryland team analyzed 1,565 papers to create a complete taxonomy of methods like Chain-of-Thought, Self-Consistency, and Tree-of-Thought. They even benchmarked techniques on MMLU to show what actually works in practice.

You should definitely read through it, then upload it to your favorite AI and have it reverse engineer all the prompting strategies and build a prompt template for you!

Treats to Try.

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

*Guidde converts your screen recordings into professional video tutorials with AI-generated step-by-step narration and voiceover in 100+ languages.
Kitten TTS is a new open source text to speech model that converts any text you write into realistic speech with multiple voice options, running locally on your device without needing a GPU (HuggingFace, GitHub)—here’s a demo.
Clay enriches your sales leads from 130+ data sources and automates your manual research (raised $100M).
Overlap clips the best moments from your hour-long podcast and auto-formats them for social posting with captions and vertical formatting.
LMArena released a new Search Leaderboard showing which models are the best at searching the web (o3-search, claude-opus-4-search and gemini-2.5-pro-grounding were the best 3).
Orchids is a design-first vibe-coding tool that builds beautiful websites and apps for you without any coding; just describe what you want and get landing pages, e-commerce sites, web apps, you name it.

Around the Horn.

Why yes, some cooky cats actually DID just hold a funeral for Claude Sonnet 3…

Google launched Guided Learning, which walks you through problems step-by-step with questions and quizzes instead of just handing you the answer.
MidJourney now creates “HD videos” with 4x more detail (at ~3.2x the price) than their standard videos, perfect when you need professional-quality footage.
OpenAI-backed Chai Discovery raised $70M at a $550M valuation in an effort to “commercialize” its new model Chai-2 to multiple pharmaceutical companies.
LangChain Labs released Open SWE, an open source cloud-based async coding tool that handles your planning, coding, and review work automatically while you do other things.
Thomson Reuters launched CoCounsel Legal, which researches legal questions step-by-step and drafts documents like privacy policies and complaints while showing its work with citations.
U.S President Trump’s Truth Social app will now provide AI search via Perplexity (in the browser version of the app).

Free Ebook: Master LLM-as-a-Judge evaluations

Learn how to evaluate your AI outputs quickly and accurately with Galileo’s in-depth eBook! Enjoy 70 pages of expert content on:

Automating 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 judges.

Get your copy now.