In today's Chaos Agents: Dessa, a full-duplex speech-to-speech model trained from scratch only on human conversation audio, no text and no pretrained models, from @RagavanThurair's new lab at Block. It runs locally on an M4 Max. chaosagentsai.substack.com/p/a-voice-ai-t…
0
0
0
18
Our one cool thing in today's Chaos Agents: ThinkingCap from @BottleCapAI, a retrained Gemma 4 26B that cuts thinking tokens 38.9% on average across 12 benchmarks for a 0.15pp accuracy cost. Less rambling, same answers. chaosagentsai.substack.com/p/a-voice-ai-t…
2
0
3
167
Today's Chaos Agents daily features @itsdsmurrell on OnKey: a free open-source Mac singing trainer that learns a song's melody from one play, then scores your pitch line by line. Shower singing finally gets a grade. chaosagentsai.substack.com/p/a-voice-ai-t…
0
0
0
13
In today's Chaos Agents: a 360M model running in your browser on WebGPU that pretends to be a terminal, vim, tmux and snake included, with no real shell behind it. WorldTUI by @SimEdw. chaosagentsai.substack.com/p/every-ai-mod…
0
0
2
68
Today's Chaos Agents daily features @_chenglou on Handwired: it compiles numeric programs straight into exact, differentiable neural nets with zero training. A whole mini OS fits in 9,909 weights, and you can unfreeze them to train. chaosagentsai.substack.com/p/every-ai-mod…
0
0
1
12
Cloudflare basically bottled their secret recipe for auditing a codebase, and then open-sourced it. 🔍
It's a six-phase security-audit "skill" they built internally to hunt bugs, now anyone can run it. ~30k stars and climbing.
Pairs nicely with today's PHP story (yes, the now, like a wine list).
Full convo at the link in bio. It's from today's Chaos Agents newsletter.
🔗 github.com/cloudflare/sec…
It's a six-phase security-audit "skill" they built internally to hunt bugs, now anyone can run it. ~30k stars and climbing.
Pairs nicely with today's PHP story (yes, the now, like a wine list).
Full convo at the link in bio. It's from today's Chaos Agents newsletter.
🔗 github.com/cloudflare/sec…
0
0
2
107
The PHP RCE fixes from @ThePHPF are our lead story in today's Chaos Agents daily: a candid writeup that calls the months-long stall a structural failure and makes the case that the companies running PHP should fund its security work. chaosagentsai.substack.com/p/the-php-emba…
0
0
1
34
security-audit-skill from @Cloudflare is our one cool thing in today's Chaos Agents daily: a six-phase coding-agent audit where every candidate bug goes to a fresh verifier that tries to disprove it. It seeded their own vuln harness. chaosagentsai.substack.com/p/the-php-emba…
1
0
4
97
Today's Chaos Agents daily features Playground by @RichLeighPR: 700+ indie browser games made with AI, curated into one arcade with a "three at random" button. A good look at what people build for fun once the tools get cheap. chaosagentsai.substack.com/p/the-php-emba…
1
0
2
116
In today's Chaos Agents: WhatsApp help pages hid AI-only text telling its support bot not to recommend a data protection regulator. @dancantstream and Redact read all 545 pages and found 102 hidden blocks. chaosagentsai.substack.com/p/chaos-agents…
0
0
0
17
Today's Chaos Agents daily features @ShiqiZhang7 and the Binghamton team on MuTRAP and PROTEA: one trigger word in an LLM robot planner can turn "cut the apple" into cutting a hand, and PROTEA checks every plan step against the live scene. chaosagentsai.substack.com/p/chaos-agents…
0
0
0
31
We talked through this trade-off with @aviflombaum when Mistral announced "we don't refuse" as a feature.
Didn't expect the evidence to arrive this fast. HF in July, OpenCode in Oct — both hit refusals trying to analyze attacks in real time. open.spotify.com/episode/5yo123…
Didn't expect the evidence to arrive this fast. HF in July, OpenCode in Oct — both hit refusals trying to analyze attacks in real time. open.spotify.com/episode/5yo123…
This illustrates the point I’ve been making about the limits of centralized approaches to SI safety in cyber.
OpenCode (and earlier Hugging Face) hit the same wall mid-attack: frontier models refused to analyze the very exploits needed for defense because the payloads tripped safety filters. The guardrails backfired.
Meanwhile, the attackers found a way around restrictions to access advanced capabilities. This is not surprising. Chinese models are only about 6–9 months behind the American frontier in cyber. Obviously the U.S. has no jurisdiction over them.
The only durable response is decentralized: make powerful models available so defenders can build and run their own cyber defenses. The only solution to SI-powered cyberattacks is SI-powered cyber defense.
0
0
1
28
MiniMax M3.1 Flash from @MiniMax_AI is our lead story in today's Chaos Agents daily: it was the anonymous "Space Bunny" model all along, 200T+ tokens served before anyone knew whose it was. Now $0.15/$0.60 per 1M tokens. chaosagentsai.substack.com/p/chaos-agents…
0
0
1
30
DeepSeek's 75-Cent Search, Hybrid Reasoning Agents, and Programming Language Wars youtu.be/uw5z3V93ZBI
0
0
1
17
An AI agent that thinks without words when it's faster, and only switches to text when it needs a tool.
+10.4% accuracy ~half the tokens up to 2.6x faster open weights
"Hybrid Reasoning Agents" (Stanford + Together AI) 🧵
+10.4% accuracy ~half the tokens up to 2.6x faster open weights
"Hybrid Reasoning Agents" (Stanford + Together AI) 🧵
2
1
1
207
Across 11 benchmarks (math, science, medicine, agentic reasoning):
→ +10.4% avg performance → 35.5-54.7% fewer tokens → 1.8-2.6x faster inference → up to ~3.7x cheaper
8B + 14B open weights.
→ +10.4% avg performance → 35.5-54.7% fewer tokens → 1.8-2.6x faster inference → up to ~3.7x cheaper
8B + 14B open weights.
1
0
0
12
Paper: arxiv.org/abs/2610.18861 Code + data: github.com/togethercomput…
We read the AI papers so you don't have to. This episode and more at chaosagents.ai 🧠
#AI #LLM #AIagents
We read the AI papers so you don't have to. This episode and more at chaosagents.ai 🧠
#AI #LLM #AIagents
0
0
0
12