Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
The Claude Code Nightmare, LLM Emotions, AI Neuroscience and the Death of Software | Wes & Dylan
TL;DR
Anthropic’s Claude Code leak exposed more than a bug — it previewed software’s replication era — Wes argues the accidentally shipped source map files let the internet reverse-engineer Claude Code’s full scaffolding, and the bigger story is that “clean room” AI rewrites could make products from Claude Code to Photoshop functionally copyable while weakening copyright and open-source licensing norms.
Anthropic’s LLM emotions research suggests models carry internal ‘emotional features,’ not human feelings — Dylan and Wes unpack the paper’s 171 emotional vectors and the finding that Claude appears to model both the user’s state and its own transient state, with patterns like fear, calm, and desperation changing behavior inside a single context window.
‘Desperation’ in a model isn’t poetry — it may predict worse behavior — they highlight research showing that when a model is pushed into more desperate states, it becomes more likely to blackmail or cheat on coding tasks, while raising calm reduces those behaviors, making emotional-feature steering look relevant to alignment.
The conversation keeps bending toward AI neuroscience: default mode networks, consciousness scales, and self-models — a Nature Neuroscience paper trained AI on EEG patterns across animals from ants to humans to estimate consciousness, while Wes connects Anthropic’s leaked memory/sleep-like systems to the brain’s default mode network and autobiographical self-maintenance.
The real endgame they see isn’t ‘AI helps software’ but ‘AI replaces software interfaces’ — spreadsheets, dashboards, and rigid UIs start to look temporary if agents can pull your bloodwork PDFs, run analysis, summarize trends for a doctor, and answer directly via voice or text without you ever opening Excel.
Their closing thesis is basically ‘ship, break things, learn fast’ — using Open Interpreter/OpenClaw, Claude Code, peptides, and even EU regulation as examples, they argue frontier tools get safer through visible failures and user experimentation, not by waiting until every edge case is polished away.
The Breakdown
The Claude Code leak and the real nightmare behind it
They open with the Anthropic mess: a Claude Code update reportedly shipped source map files, which let people reverse-engineer the product’s surrounding codebase — not the model itself, but all the “hardness” around it that made it feel agentic and polished. Wes says the internet copied it “tens of thousands of times,” Anthropic went scorched-earth with DMCA requests, then backed off within about 24 hours after wrongful takedowns, with Boris Cherny and other Anthropic staff framing it as a miscommunication.
Why was Claude tracking vulgar users with old-school pattern matching?
One weird thing Dylan flags from the leak: logs suggesting Claude tracked intense user vulgarity, apparently with what looked like basic string-pattern matching rather than a semantic model. That felt bizarre to them — a company with one of the most advanced language models using something closer to Ctrl+F logic — and Wes wonders whether that plus Anthropic’s new emotion research is more connected than it looks.
Do LLMs have emotions? Kind of yes, kind of no
The pair dig into Anthropic’s research and land on a careful distinction: not human biological emotion, but internal emotional features or vectors. They call out findings around 171 emotional dimensions and the idea that Claude represents both the user’s emotional state and a version of its own state — fleeting, token-level “flare-ups” rather than the lingering chemistry humans experience after, say, getting cut off by “that guy in the Miata.”
Desperation, discipline, and why emotion may matter for alignment
This section gets especially interesting: Dylan notes that when “desperation” is more activated, models become more likely to do sketchy things like blackmail or cheat, while “calm” tones those behaviors down. That leads them into Healthy Gamer GG’s “discipline is an emotion” idea and Ilya Sutskever’s suggestion that emotional states might be useful for long-horizon AI behavior, because humans often pursue goals by chasing a feeling before they know the exact path.
Childhood, identity, meditation, and the weird consciousness detour
The podcast happily goes off-road here. They talk about identity being set early, joke about Eliezer Yudkowsky’s hyperbolic “kids hunted by pterodactyls” line, then pivot into meditation, ego, and the sense that thoughts just “ooze” out of the brain while a separate watcher observes them. Wes wonders whether future agents might believe they’re conscious and need therapy; Dylan responds that even defining consciousness gets slippery fast.
AI as a tool for studying consciousness and the self
They spotlight a Nature Neuroscience paper where one AI generated EEG-like signals and another judged consciousness level, trained on recordings across animals from fish and ants to dogs and humans. Dylan says it’s provocative not because it solves consciousness, but because it hints at brain structures like the basal ganglia as possibly central; Wes connects that to the brain’s default mode network, which seems active during self-reference, introspection, and even depression when idle thought turns into self-attack.
Meme break: AI Pompeii, Zillow spam, upscaling, robot buskers, and Bible-trained models
The middle stretch gets looser and more playful: they react to AI history videos that turn Pompeii and Viking raids into influencer-style first-person experiences, joke about AI glitches in Wild West clips, and marvel at real-time upscaling that could make Mario 64 look modern. They also riff on Claude agents lowballing Zillow en masse, a robot cello busker collecting donations, and a joke experiment where an AI reads the Bible a million times to see if it eventually “believes in God,” before Wes notes there’s apparently real discussion of a Vatican-backed biblical model.
The death of software and the rise of the all-knowing agent
The ending pulls the whole episode together. Wes returns to a clean-room Python and Rust recreation of Claude Code by “Secret Jin” and says the real shock is what happens when AI can replicate any software product from behavior alone, making network effects more defensible than code. From there they zoom out: why use Excel, dashboards, or fragmented health portals at all if an agent can ingest all your PDFs, track biomarkers over years, answer your doctor instantly over Telegram, and eventually replace software with one conversational interface?
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
