Below the Ice.
Un solo tema, contado como se debe. Deep-dives nocturnos de The Penguin Alley.
Los titulares son la superficie — nosotros buceamos debajo. Un tema de IA cada noche, explicado desde cero para builders que cierran el día. De The Penguin Alley.
https://penguinalley.com/podcast/below-the-ice/feed.xmlApple Podcasts y Spotify — muy pronto
The Loop Debate That Closed the Fair
The AI Engineer World's Fair ended not with applause but with a fight about loops — we go under the headline to explain what agentic loops actually are, why the field is split on them, and what that means for what you build next.
Self-Evolving Agents: Real Growth, or a Good Benchmark?
AI agents are increasingly "improved" without any retraining — by evolving their own reflections, playbooks, and cheat-sheets while the model underneath stays frozen. Tonight we go under the headline to ask whether that self-evolution is real progress or just a win on the one benchmark that flattered it, and how held-out selection tries to tell the difference.
Retry Storms: The Day One Bad Loop Cost More Than a Month of Servers
A single day of AI API calls outran a month of server bills. We dive under the headline into retry storms — the old distributed-systems failure that token-metered pricing just made expensive again — and the backoff, jitter, and idempotency habits that tame it.
Speculative Decoding: The Guess-Ahead Trick Behind DSpark
We dive under tonight's top Hacker News story — DeepSeek's DSpark paper — to explain speculative decoding from first principles: how a small draft model lets a big model generate text faster without changing a single word of the output.
Two Models Are Faster Than One: Speculative Decoding and DeepSeek's DSpark
Your LLM writes one token at a time — which is exactly why it feels slow. Tonight we go under the hood of speculative decoding, the draft-and-verify trick behind DeepSeek's new DSpark paper, and ask whether 'lossless' speedups really come for free.
The Verification Horizon: When Checking the Code Gets Harder Than Writing It
Tonight we slip under a quiet line in a new paper: for AI coding agents, the old rule that checking an answer is easier than finding one has flipped — and that quietly breaks how we reward, trust, and train them. We trace why verification became the bottleneck and what to watch as the gap widens.
When the Model Can't Tell Who's Talking: Prompt Injection as Role Confusion
We dive under AI's most stubborn security problem — prompt injection, reframed as 'role confusion': why models can't reliably separate trusted instructions from untrusted data, and what it actually takes to fix it. Built from Simon Willison's writeup, the role-confusion paper, and Gray Swan's red-teaming conversation.
A Quintillion a Second: What Exascale Really Means for Science
Europe just switched on JUPITER, its first exascale supercomputer, and a wave of new AI-for-science tools landed at ISC in Hamburg this week. Tonight we go under the headline: what "exascale" actually means, what it lets scientists do, and what's real versus hype.
Stolen Sorrows: When AI Launders a Plagiarized Book
A bestselling author's book was lifted wholesale, fed through AI, and relaunched under someone else's name. Tonight we go beneath today's top headline to ask how AI turns plagiarism into a scalable business — and what, if anything, can stop it.
Can Your Research Agent Keep a Secret?
Your AI research agent reads the open web, ingests tool outputs, and chats with other agents — which means it can be quietly tricked into spilling your private data. Tonight we go under the headline on how those leaks actually happen and the emerging idea of 'deontic' runtime policies that tell an agent what it must, may, and must never do.
Eighteen Answers: AI Reasoning Meets the Rare-Disease Diagnosis
An OpenAI reasoning model just surfaced 18 new diagnoses in childhood rare-disease cases that had gone unsolved for years. Tonight we go under the headline: how these models actually reason over a child's symptoms and genes, why families wait so long for an answer, and where the hype outruns the science.
When Code Became Free — And Engineering Got Harder
In 2025 the economics of code production flipped: generating code became instant and nearly free. Tonight we go under Charity Majors' argument that cheap code makes engineering discipline more essential than ever — and what that means for how you actually build.
When the Government Pulled the Plug: Anthropic's Fable 5 Export Freeze
A calm walk through the U.S. export control directive that took Fable 5 and Mythos 5 offline for foreign nationals — what the order actually does, the personality fight Axios surfaced behind it, and what builders should watch next.
After AGI: What ASI Actually Means — and Why the Definition Wars Matter
Every major lab has AGI penciled in as a near-decade target. Two papers published today ask the question nobody wants to sit with: what exactly are we building toward, and what happens when we get there?
Música del tema: "Dreams Become Real" de Kevin MacLeod (incompetech.com), CC BY 4.0.