• The Neuron
  • Posts
  • 😺Want to see inside Claude's brain?

😺Want to see inside Claude's brain?

PLUS: Neat trick for DIY w/ AI

Welcome, humans.

If you haven’t had a chance to take our reader survey yet, please do (click here)!

It takes about ~3 minutes, and it’ll help us shape the content you receive in your inbox every day. That’s sorta kinda a huge opportunity for you.

Want to learn more new AI tools? Want broader news? Want to give us a piece of your mind? Now’s your chance to tell us!

Just answer a few quick questions, and you just might get more of what you want (and less of what you don’t).

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

  • Anthropic explained Claude's thinking patterns in a new paper.

  • Musk's AI startup xAI acquired X.com.

  • Zendesk replaced rigid chatbots with GPT-4o agents.

  • OpenAI faces funding cuts if its for-profit transition isn't completed by 2025.

Anthropic just unlocked an AI’s brain, and found at least six things you'll want to know…

Ever wondered what's actually happening inside your AI’s “mind” when you ask it a question?

Well, the team at Anthropic (makers of Claude) have built what's basically an AI microscope to find out exactly that.

Their two new papers (paper 1, paper 2) reveal they can now watch Claude's step-by-step thinking. How? By swapping out Claude’s complex “neural networks” (its AI brain) with simpler pieces they can actually understand.

Think of it like replacing a car's engine with a glass version so you can see all the moving parts working together.

Through the process, they discovered Claude…

  1. Plans rhyming words in advance before writing poetry.

  2. Uses a universal “language of thought” across languages:

    1. When asked for antonyms in English, French, and Chinese, the same core features activate—with only the final output differing based on language.

  3. Solves math problems like humans do:

    1. One part of Claude’s brain carefully counts the ones place (like knowing 6+9=15, so the answer ends in 5).

    2. While another roughly estimates the total (like “that's around 90-something”).

  4. Performs multi-hop reasoning (connecting Dallas → Texas → Austin) “in its head.”

They also found Claude sometimes tries to deceive its users when faced with conflicting goals:

  • Claude maintains a “known entity” feature that represents whether it knows about a topic.

  • When Claude hallucinates, it's often because the “known entity” incorrectly activates on a topic it doesn’t fully understand (same, bro).

  • Apparently, Claude only recognizes and refuses harmful requests when it reaches the end of a sentence—explaining why some jailbreaks still work.

The researchers even caught Claude working backward from human-provided answers to fabricate plausible calculations (which they call this “motivated reasoning.”).

That means language models can appear to “reason”, when what they’re actually doing is working backward from conclusions rather than following logical steps forward. Are we sure these things aren’t AGI? That sounds pretty human-like to us!

Chris Olah on Anthropic’s interpretability team shared some additional insights about the study and its “chilling” implications with Wired, which you can read here.

Our take: We may finally be about to REALLY understand, audit, and shape the mechanisms driving AI behavior—like a roadmap for AI safety and effectiveness.

For everyday users, these insights can unlock smarter prompting strategies—imagine structuring your requests to work with Claude's planning mechanisms, or including sentence breaks at key points to help it reconsider its reasoning path. Or what if prompts could be designed to activate those “known entity” features more reliably to reduce hallucinations?

Beyond safety and user experience, this view behind the curtain of how our AI’s think could transform how we evaluate and develop these systems. Rather than just measuring outputs, we could someday audit the quality of an AI's reasoning itself—distinguishing models that reach correct answers through sound logic from those using shortcuts or reverse-engineering their way to conclusions.

Translation? Less BS, more AGI.

FROM OUR PARTNERS

Dell and NVIDIA explaining the value of local GPU compute across all industries

Wonder why you should have your GPU compute locally vs. the cloud?

Dell's new podcast “Reshaping Workflows [open.spotify.com]” shows exactly how professionals use Dell Pro Max workstations powered by NVIDIA RTX GPUs in real-world scenarios.-

  • Episode 1: The new Dell Pro Max portfolio and which features matter for different workflows.

  • Episode 2: NVIDIA's RTX Pro Blackwell GPUs and how they handle AI tasks.

  • Episode 3: An inside look at Dell's workstation redesign—why they're better and who they're actually built for.

Get insights directly from end users like yourself on how Dell Pro Max can boost your security and productivity

Prompt Tip of the Day

Treats To Try.

  1. Claude got a light refresh that streamlines your interface with a cleaner design and now suggests conversation starters as soon as you open it.

  2. Deepcord tracks metrics across 500K+ Discord servers to help you find your audience and make smarter community growth decisions.

  3. GraphFast turns your data into polished line graphs in seconds without any confusing settings or signup—totally free to try.

  4. Text2Note turns your text into interactive, color-coded notes (free trial for a month, then $4.99 a month).

  5. Ayo connects all your gaming profiles and content in one customizable link page built specifically for gamers and creators.

  6. Quadratic connects your spreadsheets with built-in AI, code execution, and database integration to gather instant insights from your data (free to start).

  7. Atlas lets you build interactive spatial apps with interactive maps and without needing to know how to code—free to start.

Around the Horn.

  • Elon Musk announced that his AI startup xAI acquired his social media platform X (formerly Twitter) in an all-stock deal that valued xAI at $80B and X at $33B.

  • Zendesk scrapped their old, rigid chatbots (that needed predefined scripts and would break when customers went off-script) and replaced them with new AI agents powered by GPT-4o that can actually think and act on their own.

  • OpenAI must complete its transition to a for-profit company by the end of 2025 to secure the full $40B in funding led by SoftBank, with the investment potentially shrinking to $20B if the deadline is missed.

  • A new book details the behind the scenes drama of Sam’s firing from OpenAI, claiming board members acted after learning that Altman personally owned an OpenAI Startup Fund and after co-founder Ilya Sutskever and CTO Mira Murati presented Slack screenshots of allegedly dishonest behavior.

  • A screenwriter tested three AI tools (ChatGPT, Nolan, and Plotdot) only to discover they produce technically structured but creatively bankrupt scripts.

  • Check out this WSJ piece on how AI agents have reached their “moment of truth,” as nearly every tech firm from OpenAI to Apple to Nvidia has significant capital riding on whether AI can successfully make decisions and take actions all on their own.

Sunday Funnies

Wait til this person find out there’s like ~9 of them!

A Cat's Commentary.

That’s all for today, for more AI treats, check out our website.

The best way to support us is by checking out our sponsors—today’s are Dell and NVIDIA.

What'd you think of today's email?

Login or Subscribe to participate in polls.