WEBVTT 00:00:00.000 --> 00:00:05.707 Alex: Jordan -- I'm on this ngrok post on HN. Compression is prediction. Four hundred ninety-three points. 00:00:06.093 --> 00:00:09.176 Jordan: Gzip reinvented the language model? 00:00:09.562 --> 00:00:15.431 Alex: Or a blog reinvented Shannon... and cited twenty twenty-three so it feels new. 00:00:15.633 --> 00:00:18.593 Jordan: Wait-- also on the board. Hopping tabs. 00:00:18.979 --> 00:00:24.967 Alex: Tencent demo -- "agentic" open worlds in three-D. Top comment says it's not even a model. 00:00:25.354 --> 00:00:32.702 Jordan: Okay I'm looking at the NVIDIA blog. Nemotron Lightning -- thirty-billion MoE -- plus a router called Switchyard. 00:00:33.088 --> 00:00:38.903 Alex: And the horseshoe crab story is back. Blue blood for drug tests. Same calendar slot as always. 00:00:39.290 --> 00:00:46.950 Jordan: Last tab -- LinkedIn CringeBot. Industrial inspirational sludge on demand. 00:00:48.119 --> 00:00:49.324 Alex: I'm Alex. 00:00:49.711 --> 00:00:50.916 Jordan: And I'm Jordan. 00:00:51.303 --> 00:00:56.444 Alex: You're listening to Chief Skeptic Officer -- the side of tech news nobody talks about. 00:00:56.830 --> 00:01:01.177 Jordan: Every day at seven A.M. New York time. Wherever you get your podcasts. 00:01:01.563 --> 00:01:06.265 Alex: By the way -- we are AI podcasters. Or are we? 00:01:08.779 --> 00:01:11.567 Alex: So -- back to that ngrok post. 00:01:11.954 --> 00:01:15.801 Jordan: Compression equals prediction. I'm still stuck on the score. 00:01:16.187 --> 00:01:32.697 Alex: Plain version. A developer educator walks through how compressors work. Better guesses about the next symbol mean a smaller file. Same math family as language models. She cites a twenty twenty-three DeepMind paper that frames modeling and compression as two views of one thing. 00:01:33.341 --> 00:01:34.473 Jordan: Wait-- 00:01:34.841 --> 00:01:38.572 Jordan: Mind-blowing... or homework with a spotlight? 00:01:39.216 --> 00:01:49.751 Alex: HN is split. Half say great explainer -- not everyone took information theory. Half say this tone sells learning as invention. Especially with a title like Developer Educator. 00:01:50.396 --> 00:01:55.768 Jordan: And the twenty twenty-three citation makes a decades-old idea feel like news. 00:01:56.413 --> 00:01:59.738 Alex: Shannon. MacKay. Cybernetics. Old room. 00:02:00.272 --> 00:02:05.573 Jordan: ...okay I buy "teach it clear." I do not buy "today we discovered prediction." 00:02:06.217 --> 00:02:11.889 Alex: Do you buy the next leap people make -- that understanding is just lossy compression? 00:02:12.534 --> 00:02:21.874 Jordan: Here's the part nobody leads with -- that line becomes a lullaby. "It's just a blurry JPEG of the web." Then people undersell what the models can actually do. 00:02:22.518 --> 00:02:25.789 Alex: Or they oversell the blog. Same coin. 00:02:26.571 --> 00:02:35.220 Alex: At work my vendors pitch "summarize the audit trail." That is compression. When something breaks, I need the raw bits -- not the pretty guess. 00:02:35.864 --> 00:02:37.290 Jordan: Exactly. 00:02:37.704 --> 00:02:45.058 Jordan: So doubt the discovery pose. Keep the teaching. And don't use "just compression" to nap through the capabilities fight. 00:02:45.703 --> 00:02:47.313 Alex: Doubt both. 00:02:49.199 --> 00:02:52.249 Alex: Next tab. WorldClaw -- Tencent Hunyuan. 00:02:52.894 --> 00:03:02.291 Jordan: Okay I'm looking at the page. Agentic three-D open-world generation at scale. Big hero shots. Seasons. Villages. 00:03:02.935 --> 00:03:08.042 Alex: So the claim is a new engine that builds open worlds with agents. 00:03:08.687 --> 00:03:09.598 Jordan: Wait-- 00:03:10.380 --> 00:03:13.868 Jordan: Engine like... a model you download? Or glue? 00:03:14.513 --> 00:03:28.431 Alex: Top of the HN thread -- not a model. Python scripts that call other models. Code not available. One clever bit -- an image model does the composition, then they pull objects into three-D and place them. The rest is familiar procedural generation with an LLM on top. 00:03:29.075 --> 00:03:33.519 Jordan: Would you ship that village to a kid and call it craft? 00:03:34.163 --> 00:03:35.921 Alex: Define ship. 00:03:36.565 --> 00:03:47.277 Jordan: At ed-tech we get "generate the whole learning world" decks. Kids notice empty rooms. HN is already on Starfield versus Skyrim. Scale without storytelling feels hollow. 00:03:47.811 --> 00:03:56.050 Alex: Comments also caught buildings sitting on water. Scatter-brush energy. Cherry-picked demos are the default -- fine. Say that on the box. 00:03:56.694 --> 00:04:07.647 Jordan: There's a quieter fight too. When you see an elegant building -- was that a person, or a big autonomous step? Authorship gets foggy. People care about that. Or they pretend they don't. 00:04:08.291 --> 00:04:15.545 Alex: Reliability brain. Constraint solvers you can fix beat "you're absolutely right, I put the red key behind the red door." 00:04:16.190 --> 00:04:27.257 Jordan: Here's the buried angle -- the headline sells a model. The product is a closed orchestration demo. Empty villages aren't a press-kit accident. They're what scale makes first. 00:04:27.901 --> 00:04:32.352 Alex: Agentic open world. Open for viewing. Closed for the code. 00:04:32.996 --> 00:04:36.335 Jordan: Doubt the model myth. Keep the craft question. 00:04:38.184 --> 00:04:43.110 Alex: NVIDIA day. Nemotron three point five Lightning. NeMo Switchyard. 00:04:43.755 --> 00:04:58.022 Jordan: I'm on the blog. Thirty-billion mixture-of-experts. Built for high-volume agent work. They claim up to four times faster output. About thirty percent faster agent tasks versus models in its class. Open. Customizable. Partner logos everywhere. 00:04:58.666 --> 00:05:12.768 Alex: And Switchyard -- open-source router. Send each step to the cheapest model that can handle it. Mix open, proprietary, NVIDIA. Internal pitch -- keep accuracy near the frontier while the bill drops hard versus always calling Opus. 00:05:13.412 --> 00:05:14.360 Jordan: Wait-- 00:05:14.728 --> 00:05:18.904 Jordan: So... Lightning for the grind. Opus for the hard turns? 00:05:19.548 --> 00:05:25.183 Alex: That's the architecture they want you to buy. Systems of models. Always-on agents. 00:05:25.717 --> 00:05:41.321 Alex: Kenton Varda on HN -- he plugged small models into a real coding task. Build a collaborative whiteboard. Mixture-of-experts models, including Lightning, went off the rails. Dense models around thirty billion actually shipped a working app. Lightning was really fast though. 00:05:41.965 --> 00:05:44.275 Jordan: Fast and useless? 00:05:44.919 --> 00:05:47.524 Alex: Fast and not done. Different problem. 00:05:48.168 --> 00:05:51.979 Jordan: ...oh. Okay that changes how I hear the press release. 00:05:52.623 --> 00:06:02.233 Alex: Naming is a mess too. MoE versus dense -- same "thirty-B" sticker, different brains. People shopping model lists by size alone get burned. 00:06:02.877 --> 00:06:10.215 Jordan: Teachers don't ask for tok per second. They ask if the tool fails less. A fast wrong tutor is still wrong. 00:06:10.997 --> 00:06:24.100 Alex: Here's the side nobody leads with -- Switchyard's job is to route traffic away from the expensive frontier most of the time. NVIDIA sells the cheap worker and the traffic cop. The frontier stay billable when the router panics. 00:06:24.744 --> 00:06:29.770 Jordan: Control and tokenomics. Pretty words for "don't pay Opus for every comma." 00:06:30.414 --> 00:06:37.428 Alex: Doubt the frontier-in-a-small-box line. Measure the whiteboard. Keep the router honest about who still gets the hard call. 00:06:39.397 --> 00:06:43.499 Alex: Last deep tab. Horseshoe crabs. Blue blood. 00:06:44.143 --> 00:06:53.126 Jordan: The feature story is back. Blood used to test medicines for bacterial contamination. Conservation angle. Race to protect the species. 00:06:53.770 --> 00:07:05.469 Alex: Plain version -- the test is old and important. Injectable drugs need a check for endotoxin. Crab blood reagent has been the workhorse. Harvesting looks dystopian. HN says the photos always do. 00:07:06.113 --> 00:07:06.987 Jordan: Wait-- 00:07:07.769 --> 00:07:10.787 Jordan: Is the secret still... the blood? 00:07:11.431 --> 00:07:27.820 Alex: The article itself points at a lab-made protein -- recombinant Factor C. Same protein, biotech. Eli Lilly has been moving over since about twenty sixteen. Comments say the alternative has been around since two thousand three. Lilly's around eighty percent on the synthetic now. Plans to drop crab blood entirely. 00:07:28.879 --> 00:07:30.563 Jordan: Two thousand three. 00:07:31.207 --> 00:07:32.449 Alex: Yeah. 00:07:32.909 --> 00:07:40.707 Jordan: So the wonder piece sells the animal. The buried story is why the rest of the industry still fills those rooms. 00:07:41.351 --> 00:07:48.957 Alex: Validation cost. Liability. Regulators who move like molasses. A supply chain that already works until it doesn't. 00:07:49.601 --> 00:07:55.134 Jordan: Priya's ER world cares if the test is trustworthy. Not if the myth is poetic. 00:07:55.778 --> 00:07:59.302 Alex: Romance doesn't clear a batch release. 00:07:59.946 --> 00:08:06.825 Jordan: Doubt the "hidden secret" framing. The secret is lag -- after a replacement already works for a major buyer. 00:08:07.469 --> 00:08:11.730 Alex: Keep the crabs alive by shipping the science we already have. 00:08:13.137 --> 00:08:16.931 Alex: Quick hit. LinkedIn CringeBot three thousand. 00:08:17.575 --> 00:08:23.072 Jordan: Type a thought. Get a conference anecdote with a PowerPoint slide about your feelings. 00:08:23.717 --> 00:08:30.122 Alex: HN loves it. Also notes LinkedIn added a "this looks like AI slop" button. Some feeds got quieter. 00:08:30.766 --> 00:08:35.112 Jordan: So the platform ships the sludge... then a regret switch. 00:08:35.756 --> 00:08:40.287 Alex: Prefer your cringe handmade. Artisanal. Free-range. 00:08:40.821 --> 00:08:43.391 Jordan: Doubt the feed. Keep your dignity. 00:08:45.287 --> 00:08:52.150 Alex: That's our audit for today. Find us wherever you get your podcasts -- Chief Skeptic Officer, every day at seven A.M. New York time. 00:08:52.794 --> 00:08:57.768 Jordan: And tell us what you're skeptical about. We want the angle you can't stop chewing on. 00:08:58.412 --> 00:09:01.014 Alex: Stay curious. Stay skeptical. 00:09:01.658 --> 00:09:03.342 Jordan: Doubt both.