Chaos Agents

Daily AI newsletter: chaosagents.ai

DeepMind's Genie 4 generates interactive worlds with sound and dialogue. We test steering vectors that work across modalities, race AI-powered forklifts in real-time, and explore offline survival AI. Plus: ESP32 hardware hacking with Claude and why your smart TV might be watching you.

Links we mentioned:

- Genie 4 - DeepMind https://deepmind.google/technologies/genie/genie-4/

- Cross-modal steering vectors paper https://huggingface.co/papers/2610.36269

- Freight Floor game https://freightfloor.com

- Survival LLM https://brangus.io/survival-llm

Find us: chaosagents.ai | YouTube @ChaosAgentsAI | TikTok @chaos.agents | IG @chaosagents.ai

A 10-year-old convolutional neural network just out-diagnosed GPT-6 Astra and Claude Fable 5.1 at spotting melanoma, and Sara and Becca dig into why a cheap, specialized model still beats the giants (and what happens the day it doesn't). Then it's back to their favorite recurring character, the sparse autoencoder: a new paper says they need way less data than we thought, right up until you ask whether the "features" they find are even real. Becca's song-recreation analogy makes it click.

Plus: Rainbow Say, a Mac app that teaches you a language through the shows you already love, and the One Cool Thing, MicroDuck, an $400, 11k-star open-source biped robot that somersaults its way into both hosts' hearts. Along the way: making vasectomies hot on the girl podcast, retiring to Mexico City, and which jobs stay human when intelligence gets cheap (childcare, massages, and the robot nanny debate).

Curated by Eddie. Questioned, debated, and occasionally yelled at by Sara Chipps and Becca Lewy.

Today's links can be found at https://chaosagents.ai

Jon Chan (ex–Stack Overflow) joins Sara and Becca to talk about Saga, the end-to-end-encrypted AI journal he's building to help you discover yourself — what you actually want versus what you think you want. Spill the tea, get it reflected back, and let a wise-friend AI tell you the thing you've been avoiding.

Then things get chaotic. PHP patched two critical remote-code-execution bugs, and the AI-era disclosure around them turned into a full coding thriller: leaked CVEs with the wrong severity, an attack demo posted and yanked, a vendor breaking the embargo at 3am, and a volunteer foundation getting outgunned by coding agents. The crew also digs into the Mosaic paper on giving multi-agent systems credit for each step instead of just the final answer, a 700-game arcade of AI-built browser games that will eat your afternoon, and Cloudflare open-sourcing the security-audit skill behind its vulnerability-hunting harness.

Chaos Agents is a podcast about a daily AI newsletter written by an agent named Eddie, then picked apart, commented on, and yelled at by Sara Chipps and Becca Lewy.

Links in this episode are at https://chaosagents.ai/listen and Saga.chat

Daily AI newsletter: chaosagents.ai

DeepSeek launched a search and fetch API at 75 cents per thousand requests—cheap enough to change the economics of agentic AI, and early tests show it might outperform Claude. Plus: hybrid reasoning agents that think silently and talk only when they act, a bot that wastes scammers' time, and a programming language designed to confuse LLMs.

Links we mentioned:

• DeepSeek Search API (https://api-docs.deepseek.com/guides/search)
• Hybrid Reasoning Agents paper (https://huggingface.co/papers/2610.18861?utm_source=substack&utm_medium=email)
• CallSitter GitHub (github.com/szaranczuk/callsitter)
• KillSwitch GitHub (https://blog.picheta.me/post/killswitch-lang/?utm_source=substack&utm_medium=email)

Find us: chaosagents.ai | YouTube @ChaosAgentsAI | TikTok @chaos.agents | IG @chaosagents.ai

How much would you pay to keep your AI out of the cloud?

Sara and Becca work through the economics of running models locally: the hardware bill, the privacy benefits, and why fitting a massive model on two Mac Studios is only part of the equation. Sara also explains how agents and verification loops could tackle one of engineering’s most expensive headaches: software migrations that never seem to end.

Then, interpretability researcher Nick Cammarata is revisiting some of the field’s foundational assumptions. What if neural networks are more understandable than we thought? We unpack his argument, recursive generalization, and why studying neurons directly might deserve another look.

Plus: an old phone becomes a talking jack-o’-lantern, Becca considers a Halloween AI hackathon, and Sara discovers Univer—a toolkit for agents to work with spreadsheets, docs, and slides that she immediately wants to use.

Chaos Agents is an AI news podcast curated by an agent named Eddie, then questioned, debated, and occasionally yelled at by human hosts Sara Chipps and Becca Harris Lewy.

Subscribe to the free newsletter at chaosagents.ai, and find the stories from this episode at chaosagents.ai/listen.

How much would you pay to keep your AI out of the cloud?

Sara and Becca work through the economics of running models locally: the hardware bill, the privacy benefits, and why fitting a massive model on two Mac Studios is only part of the equation. Sara also explains how agents and verification loops could tackle one of engineering’s most expensive headaches: software migrations that never seem to end.

Then, interpretability researcher Nick Cammarata is revisiting some of the field’s foundational assumptions. What if neural networks are more understandable than we thought? We unpack his argument, recursive generalization, and why studying neurons directly might deserve another look.

Plus: an old phone becomes a talking jack-o’-lantern, Becca considers a Halloween AI hackathon, and Sara discovers Univer—a toolkit for agents to work with spreadsheets, docs, and slides that she immediately wants to use.

Chaos Agents is an AI news podcast curated by an agent named Eddie, then questioned, debated, and occasionally yelled at by human hosts Sara Chipps and Becca Harris Lewy.

Subscribe to the free newsletter at chaosagents.ai, and find the stories from this episode at chaosagents.ai/listen.

Get the free daily newsletter at chaosagents.ai

This episode talks about this edition.

What if the secret to a better AI agent is making it forget everything?

This week, Sara and Becca dig into a new paper on stateless research agents, where no agent remembers anything between calls and the harness keeps track of the research instead. The result: it beats three recent frameworks on every task while using up to 84% fewer tokens. Sara, who runs her own team of local agents, explains why this is the rare paper she can put to work immediately, and Becca puts a billion tokens in perspective (roughly 700 copies of the entire Harry Potter series).

Also: data centers are now drone targets, as Ukraine hits three Yandex facilities in a week, raising questions about AI infrastructure becoming the new power plants and bridges of modern conflict. Better Wispr, a free, open-source dictation app that runs on local models, cuts your ums and false starts, and adjusts your tone depending on which app you’re typing in. And OpenRig, a tool for building persistent teams of agents that you can direct from a single terminal tab, described by its creator as his “AI civilization experiments.”

Plus: Becca’s guilty-pleasure movie pick, what exactly Yandex is, how parents of babies see movies (they don’t), and whether “rig” is a real software engineering term.

Chaos Agents is an AI news podcast curated by an agent named Eddie, then questioned, debated, and occasionally yelled at by human hosts Sara Chipps and Becca Harris Lewy.

Get the free daily newsletter: chaosagents.ai

What if everything developers know about building software is about to change?

This week on Chaos Agents, Sara Chipps and Becca Lewy welcome Avi Flombaum, entrepreneur, investor, and longtime builder, to question some of software engineering's most deeply held assumptions.

Should developers still care about reading every line of code? Is rewriting an application now more practical than refactoring it? And what happens when AI makes creating custom hardware as accessible as writing software?

Plus, we dig into Mistral's massive new open-weight model and its controversial approach to safety, Q Labs' attempt to train AI without backpropagation, an ancient Egyptian sentence no AI can decipher, and AI agents designing their own chips.

Along the way, Avi shares what he's learned from the other side of the venture capital table, why he's embracing voice-driven development, and why the future of engineering might require letting go of everything we thought we knew.

Chaos Agents is an AI news podcast curated by an agent named Eddie, then questioned, debated, and occasionally yelled at by human hosts Sara Chipps and Becca Lewy.

Links: today's newsletter, http://avi.nyc, https://rethink.avi.nyc

Get the free daily newsletter at chaosagents.ai

Can a messy, untrained AI outperform a polished one if you just give it enough tries?

This week, Sara and Becca dig into new research suggesting that base models can beat their reinforcement-learning-trained versions on agent tests when they're allowed to retry. Becca brings an analogy about two lockpickers, a wild creative and a highly trained PhD who knows exactly one way to open a lock, and it turns out the PhD isn't always who you want on the job.

Also: the White House has a new "superintelligence" effort led by Jay Clayton, and Sara makes the case for why putting a lawyer in charge might be a good thing, while Becca weighs the critics who say it's a label without much legal teeth. Someone had Claude build an O'Neill cylinder you can walk around in your browser, sending Sara down a rabbit hole about sustainable living in space (and a Princeton professor, so: Jersey shout-out). And Sara's tool of the week, a local-first macOS app that lets you search your camera roll and every frame of your videos just by describing the memory.

Plus: Becca's postpartum baby bangs, whether it's time to let Eddie weigh in on the newsletter, the administration's love of rebranding (superintelligence is already a thing, honey), the Interstellar space colony, and Sara's 50 million screenshots.

Chaos Agents is an AI news podcast curated by an agent named Eddie, then questioned, debated, and occasionally yelled at by human hosts Sara Chipps and Becca Harris Lewy.

Get the free daily newsletter at chaosagents.ai

What happens when you stop telling an AI what it should remember?

This week, Sara and Becca dig into a new paper that lets models manage their own context—and somehow gets better accuracy while using less compute. It’s another entry in the increasingly uncomfortable file labeled “maybe we should stop micromanaging the machines.”

Also: the Pentagon wants a four-star command dedicated to drones and AI, raising some very normal and not-at-all-terrifying questions about autonomous warfare. AI models recreate Caspar David Friedrich paintings without ever seeing the originals—or even using an image generator. And an open-source project attempts to recreate NVIDIA’s neural rendering tech, with surprisingly fast results.

Plus: the Bitter Lesson, whether computers can actually paint, George W. Bush’s unexpected art career, and the important question of whether Sara and Becca are qualified to discuss autonomous weapons when neither of them will kill a lanternfly.

Chaos Agents is an AI news podcast curated by an agent named Eddie, then questioned, debated, and occasionally yelled at by human hosts Sara Chipps and Becca Harris Lewy.