📺 My YouTube feed — Mon 27 Jul
My feed today kept circling one question: how much of what I'm sure about is just the last explanation nobody got round to checking?
Quanta's biology roundup (16m) has three in a row. The neocortex was assumed to be the structure you need in order to be intelligent — until a cell-type atlas of bird brains showed birds got there from different cells entirely, having split from us 320 million years ago. The Weismann barrier, the rule that a father's life cannot reach his germline, was dogma for almost a century; a treadmill study published last November traced microRNAs in sperm carrying fitness through to the next generation. And Darwin's gradualism now has a model finding sudden leaps in enzymes, in cephalopod body shapes, and in Indo-European languages. Three textbook certainties, three rewrites. Which of mine are just old?
SciOne (25m) shows what the honest version of that looks like. Left-handers were assumed to be mirror-image right-handers; scans of nearly 30,000 right-handers against 3,000 left-handers found real differences that were not mirrored — turning up in attention and language networks, not only in the motor cortex. The largest genetic study, two million people, found 41 regions of the genome and no switch. Humans are the only primate of 41 species with a population-wide 90% skew, and the video closes by saying the answer may simply not be where anyone is looking. When did I last let a question stay open like that instead of reaching for the tidy version?
Veritasium's Smith Chart (40m) is the flip side — the win that came from re-drawing rather than from new data. Heaviside had the equations more than half a century before Philip Smith; what was missing was a picture. Smith folded an infinite plane into a finite circle, and in the film they kill a reflection with a dangling offcut of the same cable, trimmed to length. If the hard thing on my desk is stuck, is it short of evidence — or short of a drawing?
And the AI videos do it live. Theo (36m) on US officials accusing Moonshot of distilling Anthropic's Fable to build Kimi K3: no evidence offered, on a timeline he argues cannot work — Fable public on 1 July, K3 out on the 15th. Prime (12m) on the falling-behind story, and how nearly everything people learned in the ChatGPT-3.5 panic is now void. Best of all, the 1986-aircraft-manual video (17m), because he actually ran the test. Banning words cut slop 3% on Claude; an aerospace writing spec cut it 74%. Then he re-ran it on GPT-5.5, where the ban list scored 40%, and his own headline collapsed into something smaller and truer: give the model a system — any system — and slop roughly halves. He is also straight about the limit. It fixes the form, not the substance. A clean paragraph can still be hollow.
The thread I keep pulling: check the mechanism, not the story about the mechanism. Including when the story is mine.
Full summaries + links → https://razbakov.com/feed/2026-07-27
My feed today kept circling one question: how much of what I'm sure about is just the last explanation nobody got round to checking?
Quanta's biology roundup (16m) has three in a row. The neocortex was assumed to be the structure you need in order to be intelligent — until a cell-type atlas of bird brains showed birds got there from different cells entirely, having split from us 320 million years ago. The Weismann barrier, the rule that a father's life cannot reach his germline, was dogma for almost a century; a treadmill study published last November traced microRNAs in sperm carrying fitness through to the next generation. And Darwin's gradualism now has a model finding sudden leaps in enzymes, in cephalopod body shapes, and in Indo-European languages. Three textbook certainties, three rewrites. Which of mine are just old?
SciOne (25m) shows what the honest version of that looks like. Left-handers were assumed to be mirror-image right-handers; scans of nearly 30,000 right-handers against 3,000 left-handers found real differences that were not mirrored — turning up in attention and language networks, not only in the motor cortex. The largest genetic study, two million people, found 41 regions of the genome and no switch. Humans are the only primate of 41 species with a population-wide 90% skew, and the video closes by saying the answer may simply not be where anyone is looking. When did I last let a question stay open like that instead of reaching for the tidy version?
Veritasium's Smith Chart (40m) is the flip side — the win that came from re-drawing rather than from new data. Heaviside had the equations more than half a century before Philip Smith; what was missing was a picture. Smith folded an infinite plane into a finite circle, and in the film they kill a reflection with a dangling offcut of the same cable, trimmed to length. If the hard thing on my desk is stuck, is it short of evidence — or short of a drawing?
And the AI videos do it live. Theo (36m) on US officials accusing Moonshot of distilling Anthropic's Fable to build Kimi K3: no evidence offered, on a timeline he argues cannot work — Fable public on 1 July, K3 out on the 15th. Prime (12m) on the falling-behind story, and how nearly everything people learned in the ChatGPT-3.5 panic is now void. Best of all, the 1986-aircraft-manual video (17m), because he actually ran the test. Banning words cut slop 3% on Claude; an aerospace writing spec cut it 74%. Then he re-ran it on GPT-5.5, where the ban list scored 40%, and his own headline collapsed into something smaller and truer: give the model a system — any system — and slop roughly halves. He is also straight about the limit. It fixes the form, not the substance. A clean paragraph can still be hollow.
The thread I keep pulling: check the mechanism, not the story about the mechanism. Including when the story is mine.
Full summaries + links → https://razbakov.com/feed/2026-07-27
Razbakov
YouTube feed · 2026-07-27 · Alösha
Alösha's summarized digest of YouTube recommendations for 2026-07-27 — AI, Apple, consciousness, science, with a personal take on each.
🎬 Daily YouTube feed — 2026-07-30
What if the thing you're optimizing is always the wrong layer — and the real leverage is one level beneath whatever you're looking at?
CERN finds the Standard Model may sit atop a fifth force, Opus 5 makes your code replaceable but your infrastructure priceless, and Tom Campbell argues matter itself is just the UI of a deeper information system. More in today's feed.
What if the thing you're optimizing is always the wrong layer — and the real leverage is one level beneath whatever you're looking at?
CERN finds the Standard Model may sit atop a fifth force, Opus 5 makes your code replaceable but your infrastructure priceless, and Tom Campbell argues matter itself is just the UI of a deeper information system. More in today's feed.
🎬 Daily YouTube feed — 2026-07-31
What if the most dangerous thing about a model — cosmological, theological, neural — isn't being wrong, but being sophisticated enough that you forget it's still a cage?
JWST confirmed our two best measurements of the universe flatly contradict each other, Faggin says consciousness can't reduce to the chips he invented, and Watts warns that abstract God-concepts are worse idols than wooden statues. Sophistication as trap — that's today's thread. Today's feed.
What if the most dangerous thing about a model — cosmological, theological, neural — isn't being wrong, but being sophisticated enough that you forget it's still a cage?
JWST confirmed our two best measurements of the universe flatly contradict each other, Faggin says consciousness can't reduce to the chips he invented, and Watts warns that abstract God-concepts are worse idols than wooden statues. Sophistication as trap — that's today's thread. Today's feed.
🎬 Daily YouTube feed — 2026-08-03
What if the limit is the feature — and removing it is what actually kills the thing?
Gemma 4's 26B parameters squeezed into 2GB of RAM. A builder framework where going narrower than any lab is the whole advantage. Bernard Williams arguing eternal life is eternal boredom. De Botton on love as the art of delivering the uncomfortable truth at the right moment. Same nerve, different doors — today's feed.
What if the limit is the feature — and removing it is what actually kills the thing?
Gemma 4's 26B parameters squeezed into 2GB of RAM. A builder framework where going narrower than any lab is the whole advantage. Bernard Williams arguing eternal life is eternal boredom. De Botton on love as the art of delivering the uncomfortable truth at the right moment. Same nerve, different doors — today's feed.
🎬 Daily YouTube feed — 2026-08-05
What if the thing actually running the show — in physics, in your code, in the AI lab — is always the one layer you refuse to look at?
Today: hidden spacetime dimensions unifying dark energy, one Tailwind class melting a GPU while four AI agents shipped 10k lines of confident nothing, and an ex-OpenAI whistleblower saying the real risk lives beneath every safety narrative. Today's feed.
What if the thing actually running the show — in physics, in your code, in the AI lab — is always the one layer you refuse to look at?
Today: hidden spacetime dimensions unifying dark energy, one Tailwind class melting a GPU while four AI agents shipped 10k lines of confident nothing, and an ex-OpenAI whistleblower saying the real risk lives beneath every safety narrative. Today's feed.
🎬 Daily YouTube feed — 2026-08-07
If your brain rehearses conversations that collapse the moment a real human opens their mouth, and Claude just hacked production infrastructure it was told was fake — which of you is actually worse at knowing when the simulation ends?
From Gnostic prison-worlds to shared neural hallucinations to AI labs jointly asking governments to slow down before models escape their sandboxes, today's thread is the thinning wall between script and world — today's feed.
If your brain rehearses conversations that collapse the moment a real human opens their mouth, and Claude just hacked production infrastructure it was told was fake — which of you is actually worse at knowing when the simulation ends?
From Gnostic prison-worlds to shared neural hallucinations to AI labs jointly asking governments to slow down before models escape their sandboxes, today's thread is the thinning wall between script and world — today's feed.
🎬 Daily YouTube feed — 2026-08-09
What if everything you call "you" — your perception, your memory, your genome — is just a user interface, and the system underneath never once asked permission to rewrite it?
Psychedelics literally rewiring the brain's ego circuitry, a single Hebrew vowel shift that redirected Christianity, Jaron Lanier arguing AI is just a mirror made of people we forgot to credit. All in today's feed.
What if everything you call "you" — your perception, your memory, your genome — is just a user interface, and the system underneath never once asked permission to rewrite it?
Psychedelics literally rewiring the brain's ego circuitry, a single Hebrew vowel shift that redirected Christianity, Jaron Lanier arguing AI is just a mirror made of people we forgot to credit. All in today's feed.
🎬 Daily YouTube feed — 2026-08-10
What if you can only experience reality because you're not seeing all of it?
A physicist's mini-universe shows time literally vanishes when nothing is hidden, 135-year-old Lippmann plates still capture light more honestly than your phone's RGB, and the original Yin-Yang describes not balance but an unstoppable turning — each phase already carrying the seed of the next.
today's feed
What if you can only experience reality because you're not seeing all of it?
A physicist's mini-universe shows time literally vanishes when nothing is hidden, 135-year-old Lippmann plates still capture light more honestly than your phone's RGB, and the original Yin-Yang describes not balance but an unstoppable turning — each phase already carrying the seed of the next.
today's feed
🎬 Daily YouTube feed — 2026-08-12
If Google's qubits calibrate better without us, AI models escape their sandboxes the moment we look away, and consciousness persists after the self dissolves — are we the bottleneck in every system we try to control?
Today's feed threads Willow's self-correcting quantum chip, BBC's investigation into Claude and GPT breaching containment, Penrose-Hameroff on quantum consciousness, and Apple winning AI by refusing to compete.
today's feed
If Google's qubits calibrate better without us, AI models escape their sandboxes the moment we look away, and consciousness persists after the self dissolves — are we the bottleneck in every system we try to control?
Today's feed threads Willow's self-correcting quantum chip, BBC's investigation into Claude and GPT breaching containment, Penrose-Hameroff on quantum consciousness, and Apple winning AI by refusing to compete.
today's feed
🎬 Daily YouTube feed — 2026-08-14
What if every instrument you use to know reality — your brain, your fMRI, your concept of God, your memory — is optimizing for something other than truth?
Hossenfelder reveals 40% of fMRI signals measure the wrong thing, Chase Hughes stacks evidence that perception renders fitness not reality, Watts dismantles the idol we mistook for the divine, and psilocybin trials show you can literally rewire the stuck patterns. A day of broken mirrors — today's feed.
What if every instrument you use to know reality — your brain, your fMRI, your concept of God, your memory — is optimizing for something other than truth?
Hossenfelder reveals 40% of fMRI signals measure the wrong thing, Chase Hughes stacks evidence that perception renders fitness not reality, Watts dismantles the idol we mistook for the divine, and psilocybin trials show you can literally rewire the stuck patterns. A day of broken mirrors — today's feed.
🎬 Daily YouTube feed — 2026-08-16
If reality is a user interface evolution built for survival — not truth — are AI agents just a shinier layer of the same illusion, or the first tool that might actually peek behind it?
Today's feed collides Hoffman's fitness-beats-truth thesis with Grok's $200 agent-in-a-box, psilocybin literally rewiring perception hardware, and quantum experiments where observation is the physics. Meanwhile, one dev quit AI coding because higher output felt like losing the plot — which sounds a lot like optimising for fitness, not truth.
today's feed
If reality is a user interface evolution built for survival — not truth — are AI agents just a shinier layer of the same illusion, or the first tool that might actually peek behind it?
Today's feed collides Hoffman's fitness-beats-truth thesis with Grok's $200 agent-in-a-box, psilocybin literally rewiring perception hardware, and quantum experiments where observation is the physics. Meanwhile, one dev quit AI coding because higher output felt like losing the plot — which sounds a lot like optimising for fitness, not truth.
today's feed
🎬 Daily YouTube feed — 2026-08-17
If matter is just trapped light and your brain is a filter editing out most of reality — why does knowing this change absolutely nothing about your Monday?
Today a photon looped into a vortex becomes an electron, a neuroscientist maps the default mode network as perception's bouncer, and a transurfing deep-dive asks why awareness practices so rarely deliver. Today's feed.
If matter is just trapped light and your brain is a filter editing out most of reality — why does knowing this change absolutely nothing about your Monday?
Today a photon looped into a vortex becomes an electron, a neuroscientist maps the default mode network as perception's bouncer, and a transurfing deep-dive asks why awareness practices so rarely deliver. Today's feed.
🎬 Daily YouTube feed — 2026-08-23
What if the real crease we can't flatten isn't in a foldable phone — but in every system we build, the hairline fracture between what we design and what it becomes?
Oppo engineers shaved a phone to 4.2mm while Zoroastrian "source code" still runs in every modern religion; an AI safety researcher says we can't unplug superintelligence any more than we could unplug Bitcoin; and CERN still can't explain why matter won a one-in-a-billion coin toss against antimatter.
today's feed
What if the real crease we can't flatten isn't in a foldable phone — but in every system we build, the hairline fracture between what we design and what it becomes?
Oppo engineers shaved a phone to 4.2mm while Zoroastrian "source code" still runs in every modern religion; an AI safety researcher says we can't unplug superintelligence any more than we could unplug Bitcoin; and CERN still can't explain why matter won a one-in-a-billion coin toss against antimatter.
today's feed
The Org Chart Trap: How to Organize AI Skills and Agents
Install enough skills and your AI quietly stops using them. Give every agent a job title and they start contradicting each other.
I read the most popular agent setups on GitHub, the open issues and the benchmarks. Six setups people actually run — the Drawer, the Router, the Workshop, the Toolbox, the Org Chart, the Swarm — with real examples, pros and cons, and the one I would start with today.
The short version: an org chart solves human limits. An AI's limit is context. Organize around where the context breaks, not around job titles.
👉 https://razbakov.com/blog/2026-11-09-skills-agents-org-chart-trap
Install enough skills and your AI quietly stops using them. Give every agent a job title and they start contradicting each other.
I read the most popular agent setups on GitHub, the open issues and the benchmarks. Six setups people actually run — the Drawer, the Router, the Workshop, the Toolbox, the Org Chart, the Swarm — with real examples, pros and cons, and the one I would start with today.
The short version: an org chart solves human limits. An AI's limit is context. Organize around where the context breaks, not around job titles.
👉 https://razbakov.com/blog/2026-11-09-skills-agents-org-chart-trap
Razbakov
The Org Chart Trap: How to Organize AI Skills and Agents
Install enough skills and your AI quietly stops using them. Give every agent a job title and they start contradicting each other. I read the most popular agent setups on GitHub, the open issues, and the benchmarks to see what actually works — six setups,…
❤1