Podbit · Lex Fridman Podcast
Compilers cannot auto-vectorize anywhere close to what an expert human can do by hand. The dav1d AV1 decoder proves it: 80% assembly, 62x faster than C on pixel conversion functions, running on 3 billion devices. Moore's Law is over — squeezing hardware through hand-crafted assembly is how you get more without buying more.
NV Energy is abandoning 50,000 residents in the Lake Tahoe area after May 2027, redirecting electricity to data centers in North Las Vegas. Data centers are reportedly lying about how much power they'll consume, and the people paying the price live in low-income, low-population areas.
WiFi signals bounce off walls and anything that interferes with the signal can be mapped — meaning the military and potentially others can see outlines of objects and people in a room. The guys wonder if this means someone can see their dicks.
LLMs produce clunky writing because the internet contains far more bad writing than good writing. A prize-winning AI-generated story was traced back to Naruto fan fiction pages and random horror blogs — because that's what constitutes the bulk of the training data, not Moby Dick.
A Commonwealth Short Story Prize-winning story was almost certainly AI-generated. The editor responded with a shrug. The phrases in the winning story trace back to Naruto fan fiction and random horror blogs — because that's what LLMs are really made of.
When Intel offered to help Apple port macOS to x86, Jobs dropped a bombshell: 'I've been working on that for the last 4 releases.' Jobs had been quietly preparing Apple's core technology for a potential future switch for years before anyone at Intel suspected. This foresight — building optionality in secret — is what made Apple Silicon inevitable.
At Intel's CPU peak, the company literally laughed at NVIDIA's machines. GPUs were for gamers, not serious compute. But Jensen Huang kept building a deeper software stack — CUDA, SIMD, multi-threading — until Japanese HPC researchers realized these 'graphics cards' could tackle the world's hardest workloads. Intel even had its own answer, Project Larrabee, killed the week Gelsinger left.
TSMC's original vision — become the factory for the entire semiconductor industry — seemed so niche that Intel didn't take it seriously. Intel was vertically integrated and proprietary; TSMC standardized everything and welcomed any customer. Steady progress and a demanding Apple drove them from irrelevant to producing 5x Intel's wafer volume by 2021, now 7x.
The AI buildout has a natural circuit breaker: energy. Nobody builds data centers without power, and global energy capacity is expanding at just 4–5% per year. That constraint limits how overheated the market can get. Meanwhile the value of a token — a unit of intelligence — is nearly infinite if it improves logistics, finance, and labor. Gelsinger sees two decades of growth, not years.
Quantum has been '5 years away' for 25 years — but Gelsinger says this time is different. We now know how to build qubits, error-correct them, and write algorithms against them. Multiple modalities (trapped ions, photonic, spin) are all showing results. Engineering scale is the only remaining challenge, and he expects quantum supremacy across multiple industries before 2030, with encryption-breaking Q-Day around 2032–2033.
We use essential and analytics cookies to run Vuci. To understand how the site is used: Privacy Policy.
Install Vuci on your phone
Add it to your home screen for a faster, app-like experience.
Install Vuci on your phone
Tap the Share button, then “Add to Home Screen”.
A new version is available
Reload to get the latest Vuci.