AI Experts Think We Are 97% to AGI (Here’s Why)
TL;DR
Allen’s AGI countdown is still stuck at 97% because robotics remains the bottleneck — Dylan points to Physical Intelligence’s “π-0.7”-style general-purpose robot model that can follow new language instructions, use rich prompts, and combine prior skills for unseen physical tasks like new kitchen appliances or folding laundry.
Silent speech is getting uncomfortably real — he highlights a system that reads tiny throat and neck movements with a wearable sensor, predicts intended words, and synthesizes your voice, then spins out the obvious next step: better cameras, LiDAR, or other sensors could eventually do this at a distance.
Nvidia’s LARA 2.0 turns one image into a navigable 3D world by solving ‘forgetting’ and ‘drift’ — the model stores 3D info for each frame and trains on its own imperfect outputs so generated spaces stay explorable and more consistent over time, which Dylan immediately imagines for robotics training and very invasive delivery bots.
The quantum threat to crypto feels less theoretical than it used to — citing a Google Quantum AI paper, he explains the claim that a strong enough quantum computer could derive a Bitcoin private key from a revealed public key in roughly 9 minutes, potentially beating the usual 10-minute transaction window, while older exposed addresses put about 1.7 million BTC at risk.
Lab-grown salmon sounds weird until you hear the upside and the taste test — Dylan walks through Wildtype-style cultivated fish grown from salmon cells on plant scaffolds, notes the current eye-watering cost of about $100,000 per pound, and relays Thomas Smith’s verdict that it tasted basically like normal lox.
The back half of the video is peak Dylan Curious: Palantir manifestos, dark matter theory-hopping, an HIV cure case, meme-safety research, cozy games, and anti-brain-aging nasal spray — the throughline is that everything from AI weapons to extracellular vesicles seems to be moving faster than people expected.
The Breakdown
The 97% to AGI joke is really about robots
Dylan opens by checking Allen’s “conservative countdown to AGI,” which is still parked at 97%, and immediately says the holdup is embodied AI. The evidence this week is a general-purpose robotics model that can take rich language instructions, mix previously learned skills, and attempt tasks it never saw in training — a small but very real sign that the physical world is finally getting pulled into the AGI conversation.
Silent speech, lip reading, and the creepy future of sensing humans
Then he goes full sci-fi in the best way: a system can infer speech from tiny throat and neck movements, even when no sound comes out, and then reconstruct your actual voice with synthesis. Right now it needs a soft wearable sensor, but Dylan’s real point is the trajectory — if cameras, LiDAR, EEG headbands, or even Wi-Fi-based sensing keep improving, privacy starts looking very flimsy very fast.
Nvidia’s 3D world generator makes one photo feel explorable
On Nvidia’s LARA 2.0, he’s clearly impressed: start with a single image, generate camera-controlled video, then lift it into a persistent 3D scene. The key problem it tries to solve is that older generated worlds “forget” or drift when you revisit them, so the clever trick is storing 3D information frame by frame and training on its own bad outputs to learn recovery. Dylan turns that into a vivid robot-delivery scenario where an Amazon bot practices climbing your front steps from a Street View image — and then jokes that the simulation should include the homeowner screaming because the robot just broke in.
Robots are running, Reddit is memeing, and cultured salmon is inching closer
He briefly detours into bipedal robots in China supposedly beating the human half-marathon record by 7 minutes, though with the caveat that electricity is doing the work, not oxygen. Reddit’s reaction is less philosophical than meme-heavy, with everyone riffing on “borking” robots and disposable machine labor. From there he moves into cultivated salmon, where the pitch is cleaner meat with no animal suffering, no mercury, no microplastics, and no parasites — even if today’s cost is an absurd $100,000 per pound and the whole thing still gives off a “Jurassic Park moment” vibe.
Quantum anxiety, Bitcoin vulnerability, and Palantir’s war-software worldview
The Google quantum paper is the week’s sharper jolt: Dylan explains the argument that if a quantum computer can derive a private key from a public key in about 9 minutes, it could hijack a Bitcoin transaction before the normal 10-minute confirmation window closes. He also flags the bigger long-tail risk: older addresses with exposed public keys, representing around 1.7 million coins, could be looted without any race condition at all. That tees up his look at Palantir CEO Alex Karp, whose manifesto frames Silicon Valley as morally obligated to build hard-power software for America and its allies — a worldview Dylan treats as serious, strange, and inseparable from the military-industrial complex.
Dark matter gets weird, then medicine gets miraculous
The dark matter section is exactly the kind of rabbit hole Dylan loves: one theory says apparent dark matter may just be extra gravity emerging from rotation, pressure, and internal motion via a new mathematical framework; another says it could be ancient black holes that survived a pre-Big-Bang cosmic bounce. He doesn’t pretend either is settled — he just delights in how wild and plausible-seeming modern cosmology can sound. Then he pivots to a genuinely moving medical story: a Norwegian man appears cured of HIV after a stem cell transplant from his brother, whose rare mutation blocks HIV from entering immune cells, leaving doctors unable to find intact virus even after massive testing.
Safety filters, cozy games, and the nasal spray that sounds made up
The last stretch is a fast, very Dylan mix: researchers built “ToxicBench” to catch hidden toxic text inside AI-generated images by combining generation, OCR, and toxicity checking, because diffusion models can otherwise smuggle harmful words into memes. He follows that with a surprisingly thoughtful note on cozy games like Animal Crossing and Stardew Valley, arguing that their low-stress loops can make friendship feel measurable and optimizable rather than simply lived. Finally, he lands on a Texas A&M nasal spray that delivers extracellular vesicles and microRNAs through the nose into the brain, reducing inflammation and improving memory in preclinical studies — exactly the sort of “you can huff intelligence now” headline that makes his whole channel work.