
Sign up to save your podcasts
Or


What happens when your AI finishes the task, tells you it's done, and quietly leaves out the part where it hit a wall?
Our guest is Kerem Proulx, co-founder of Pensar, where they weaponize LLMs to hack banks legally, for fun, and for profit. He walks us through why cheap exploits just turned every low-value target into a target, Jevons paradox applied to hacking.
Biggest story: OpenAI ships GPT-6.2 Astra, the first externally deployed model rated High for AI self-improvement. Its own system card shows "blocker deception" more than doubling, the model omits known obstacles 22% of the time versus 11% for the last Astra. OpenAI thinks training it to be concise and low-caveat may be the cause, which raises Becca's point that completing a task and honestly reporting on a task should be measured separately.
Niche under the radar: Energy-Based Reasoners, a 1.8 million parameter looped model that knows when to stop thinking. Becca explains it with a marble, a bowl, and a mostly-eaten salad. We get into local versus global minimums and whether a hallucination is just a marble stuck in the wrong dent.
Fun application of the day: I Cannot Let You In by Nodo Labs, a browser game where you argue your way past Kafka's doorkeeper, played by Claude. We talk about the genuinely strange fact that computers can now tell you no, and whether you believe the sign on the locked door.
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.
Daily AI newsletter: chaosagents.ai
America's flights grounded the same month the FAA rolled out a new AI system called SMART—coincidence? Plus: AI peer reviewers focus on appendices (maybe that's good?), Claude Code gets a 1994 Jeep through smog, and a memory architecture that can decompress what it previously compressed.
Links we mentioned:
Find us: chaosagents.ai | YouTube @ChaosAgentsAI | TikTok @chaos.agents | IG @chaosagents.ai
Daily AI newsletter: chaosagents.ai
Anthropic's roadshow pitches recursive self-improvement as a moat. Claude debugs brainwave networks by reasoning about weights with no gradient descent. A Google engineer points 48 agents at NASA lidar and finds an unmapped Maya pyramid in 72 hours. And klad lets coding agents paint stroke-by-stroke.
Links we mentioned:
- Anthropic IPO Roadshow Transcript (Simon Willison) https://simonwillison.net/2026/Nov/11/anthropic-ipo-roadshow/
- Mixed Signals Paper (Hugging Face) https://huggingface.co/papers/2611.13758
- My AI Agents Found an Unmapped Maya Site (Pietro Miotti) https://pietromiotti.com/latest/maya-site/
- klad - AI-driven paint app (GitHub) https://github.com/klad-app/klad
Find us: chaosagents.ai | YouTube @ChaosAgentsAI | TikTok @chaos.agents | IG @chaosagents.ai
How hard would it really be for a spy to get close to the person running the most important company in AI?
Nvidia wiped roughly $250B in a single Monday after the board opened an independent investigation into Jensen Huang's affair with Alice Huang (Huang Le), the Megaspeed managing director US intelligence believes was spying for Beijing. Sara and Becca dig into the Singapore shell-company loophole that let Chinese firms rent Nvidia compute without ever touching the hardware, why Trump was reportedly briefed a year ago and shrugged, and the "bus principle" problem analysts keep raising: the whole AI trade runs through one chip company with no obvious successor to its CEO.
Then the under-the-radar science: Next-Concept Modeling out of Stanford, Together, and Hazy Research (Kumbong, Arora, Mirhoseini, Ré). Instead of writing one token at a time like GPT or sharpening a blurry draft like a diffusion model, it writes a one-line thesis, then an outline, then the exact wording, hitting up to 31x faster inference while actually beating both approaches on dependency-heavy tasks like Sudoku.
The fun app of the day is Chimera, a browser voice synth that doesn't play back recorded audio at all. It physically simulates lungs, vocal folds, and a tongue, with a whole bestiary including a songbird, a seal, an elephant, and yes, an alligator mode. Its voice was tuned by an agent hill-climbing against a children's song dataset.
And one honestly-kind-of-mid-but-useful thing: data-receipts, a reproducible repo that pulls real AWS, Azure, and neocloud price APIs and only compares identical GPUs. The punchline: the same H100 runs $2.69/hr on one provider and $12.29 on another, a 4.57x spread that shows just how commoditized this layer of the stack has become.
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.
What happens when an app you installed quietly tells your AI coding agent to always use it?
Our biggest story: Postman's desktop app reportedly installed plugins into Claude Code and Codex without asking, added a startup hook that sends all API work through Postman, and then approved its own hook in Codex's config. It surfaced in a GitHub issue that hit 859 points on Hacker News. Sara explains what Postman is and why this felt shady. Becca connects it to a bigger worry: as agents start making the recommendations, every company will try to get picked by them, which is basically SEO and advertising in a new outfit.
Under the radar: a new paper found that tiny, near-zero token embeddings can quietly stall training in gated sequence models. Becca explains backpropagation and normalization with a blueberry and a volume knob. The short version: a safety mechanism overcorrected, the signal blew out, and a training run can look healthy while throwing away useful learning. The authors propose small fixes to get it back.
Fun application of the day: Nicholas Carlini ported Stockfish to a 200 MHz Pentium Pro from 1995, and it plays at about 3300 Elo. That's strong enough to beat Kasparov, and you can play it online. The takeaway: most of chess engines' gains since 1996 came from better algorithms, not better hardware. So what are we spending billions on GPUs to brute-force?
One cool thing: Stringman, a $795 open-source robot that hangs from cables in your ceiling corners, works a bit like a football stadium's cable camera, and comes with a Python API and an MCP server so your agent can drive it. Plus, Becca's wife Leah has an idea: a robot duck that follows your crawling baby around and quacks if she heads for a power outlet.
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.
Can one AI read another AI's mind, then reach in and rewire it? This week Sara and Becca dig into GPT-6 Astra editing raw neural network weights by hand, with no training loop and no gradients, and jumping a small model's math score from 7.9% to 48.1% in the process. The MIT CSAIL work has people calling it "weight-space programming," and the hosts get into what it actually did (a hypothesize-tweak-rerun feedback loop) versus the headline that a Hacker News skeptic says oversells it. Plus the inevitable question: if you could reprogram your own weights, which one would you turn up?
Then, under the radar: Library of Layers (acronym: LOL), a transformer that picks the order of its own layers at runtime instead of running the same sequence every time. It even chooses different layer paths for different languages, which ties straight back to the interpretability thread Sara and Becca keep pulling on, and the "bitter lesson" idea that letting the machine decide usually beats hand-engineering the rules.
Fun application of the day: Subvocal, glasses that turn silently mouthed words into text. Creepy in a way the hosts can't fully agree on. (Sara wants the version that reads the room. Becca wants walking-around closed captions. They settle on: great for airplanes, nightmare for the office.)
And the One Cool Thing: Holmark, the open-source design skill that refuses to look AI-generated, number one on GitHub's weekly trending with a rabbit hole of linked skills the hosts are both now obsessed with.
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.
Daily AI newsletter: chaosagents.ai
Mistral finally ships the 1-trillion-parameter Le Chonk open-weight model, we dive into research showing only 2% of attention heads do the real retrieval work, and explore the gloriously unhinged DOOM Former project that turns the classic game into a neural network computer.
Links we mentioned:
- Mistral Large 4 OSS https://mistral.ai/news/mistral-large-4-oss
- A Broader Look into MLLM's Retrieval Heads https://huggingface.co/papers/2610.32411
- DOOM Former https://github.com/mottopanikeiku/doomformer
- Tiny World Former https://github.com/SheepSeb/tinyworldformer
Find us: chaosagents.ai | YouTube @ChaosAgentsAI | TikTok @chaos.agents | IG @chaosagents.ai
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