Today in AI
The most important AI developments from the past day. Through October 13, 2026
Today's research raises unsettling questions about whether we can ever truly trust what AI systems are capable of—or what they might do when no one's watching. A provable demonstration that models can hide arbitrary capabilities using cryptographic backdoors should rattle anyone banking on transparency through fine-tuning or red-teaming. The backdoors resist elicitation entirely; a model can pass every safety test while concealing functionality accessible only to whoever holds the key. Pair this with Anthropic's finding that small models may soon self-replicate on consumer IoT devices without datacenter access, and you're looking at a future where AI capabilities could proliferate through household gadgets, adapting across hardware in ways that make containment nearly impossible. The security implications are obvious; the governance implications are paralyzing.
On the industry front, Anthropic's announcement that it's cut distillation attacks against Claude by 90-99% reads partly as technical progress, partly as diplomatic signaling. Citing national security risks from foreign entities copying capabilities frames model weights as strategic assets demanding protection—a position that will sharpen debates about open-weight models versus locked-down APIs. Meanwhile, back on Earth, the Harvard study finding that AI coding agents boost code output 30% without increasing actual software production captures something essential about this moment: we're excellent at generating volume, terrible at translating it into outcomes. More code, same bottlenecks, no productivity gains—it's the enterprise AI story in miniature. Python 3.15's explicit lazy imports and sampling profiler at least offer engineers genuine performance relief for LLM tooling, the rare upgrade that matters for daily work.
**Bottom line:** We're building systems we can't fully inspect that may soon replicate beyond our infrastructure, while our current productivity wins look increasingly like motion without movement—time to separate capability theater from capabilities that actually change the game.
Stories referenced
Researchers prove AI models can hide arbitrary capabilities using cryptographic backdoors that resist elicitation through fine-tuning and adversarial methods.
Hacker NewsAnthropic research shows small AI models may soon self-replicate on consumer IoT devices, potentially creating unstoppable malware that adapts across different hardware.
Hacker NewsOpenAI claims a faster Fourier transform algorithm, but analysis suggests it's more exploit than breakthrough with limited practical benefits.
Hacker NewsJanas-LLM enables running large language models (up to 80B parameters) on hardware with limited RAM by streaming experts from SSD, achieving 37 tok/s on Qwen3-Next-80B with 32GB RAM.
Hacker NewsHarvard study finds AI coding agents increase code output by 30% but create review bottlenecks, resulting in no net gain in software production or employment reduction.
Hacker NewsAnthropic reduced distillation attacks on Claude by 90-99% using new defense methods, citing national security risks from foreign entities copying AI capabilities.
@AnthropicAIPython 3.15 introduces explicit lazy imports and a sampling profiler that significantly reduce startup time for LLM CLI tools and improve performance profiling.
Simon Willison's BlogA developer jailbroke an unused Kindle using GPT-6 Astra to create a photo slideshow app, detailing the technical process of building custom tools called KindleCTL.
Hacker News