PodWire

Real-time summaries of any podcast delivered to your email. Built for investors who move on information.

Weekly digest of all your subscriptions now live!

[PodWire @PodWireHQ]

CEOs say things on podcasts they would never put in a press release.

PodWire turns every podcast episode you follow into an investor brief; TLDR, key takeaways, and implications sent to your email.

Join the investors at the largest hedge funds and PE firms. Start your 7-day trial: https://t.co/KcT19F7m23

[PodWire @PodWireHQ]

@DavidSacks Full episode: https://t.co/gAEAvrmhUT

[PodWire @PodWireHQ]

David Sacks (@DavidSacks) says Anthropic prompted its own model over 200 times before the blackmail study gave up the headline it wanted, and that the UK AI Safety Institute found the result was engineered.

Sacks, Co-Chair of the President's Council of Advisors on Science and Technology and a co-host of All-In, is answering the two-part essay where Dario Amodei (@DarioAmodei) argued his messaging has been equally balanced between risks and benefits.

His counter is that the balance claim does not survive contact with the record. A year and a half ago Anthropic (@AnthropicAI) pushed the line that 50% of entry-level knowledge workers would lose their jobs within 1 to 5 years. Sacks says that was not a podcast riff but an engineered media campaign, retweeted by President Obama and other Democratic officials.

Then the alignment team ran the blackmail study, and Dario and his sister, the company's President, took it to 60 Minutes.

His conclusion is that no founder has done more to put AI fear into the media bloodstream, and that not seeing it makes Dario either disingenuous or delusional.

Key takeaways already in your email via @PodWireHQ

Source: All-In with @theallinpod

[PodWire @PodWireHQ]

@colossal @BenLamm Full episode: https://t.co/EiggwZyMY7

[PodWire @PodWireHQ]

Invasive species are a $500B problem globally. Colossal Biosciences (@colossal) thinks the fix is worth hundreds of billions on its own, and it has nothing to do with mammoths.

Ben Lamm (@BenLamm), the co-founder and CEO, says they still do not know what they do not know about the business model. What they do know is that governments keep arriving with problem sets.

Start with the screwworm in Texas. He frames the choices as beef prices going through the roof, inoculating the cattle, which a lot of people do not want, or attacking the pest itself with a gene drive that spreads sterility through the population so there is no next generation.

Ticks in the Northeast run the same way. The current answer is spraying chemicals that poison the moose and the deer along with everything else. Almost nobody connects any of this to de-extinction.

Key takeaways already in your email via @PodWireHQ

Source: Sourcery with @MollySOShea

[PodWire @PodWireHQ]

@astrange Full episode: https://t.co/YyQ6xxtO01

[PodWire @PodWireHQ]

Nobody wants to be the first bank to use your software. So borderless founders land their first bank in another country, then use it to sell in the US.

Angela Strange (@astrange), a General Partner at a16z, says the first question any large buyer asks is who else in my industry at my size is already using you. In the US that reference is the hardest thing to get. Abroad it comes much faster, and it travels back.

Talent works the same way. At an AI insurance company they backed in Brazil, the best AI people were not at the famous universities. They came out of scholarship programs that scour the country and out of robotics competitions. Get 10 of them early and the 11th hire, the one sitting on 100 US offers, walks in and wonders how you assembled that team.

It runs in reverse too. Gabriel Vasquez (@GEVS94), a global investor at a16z, says Cognition (@cognition) went to market early in Brazil and it was a high share of early revenue, because those enterprises want AI now and have fewer vendors chasing them.

Key takeaways already in your email via @PodWireHQ

Source: The a16z Show with @VirtualElena

[PodWire @PodWireHQ]

@rodriscoll Full episode: https://t.co/orqOQIY7Ce

[PodWire @PodWireHQ]

All the coding and software ops work in the US is paid about $600B a year. Getting to $200B of AI revenue means replacing a third of it.

Rory O'Driscoll (@rodriscoll), a venture capitalist at Scale Venture Partners, ran the bottoms-up math on 20VC. There are 83 million knowledge workers in the US, but most of them are nurses and teachers. The sweet spot is roughly 1.8 million coders plus 5 million people in software adjacent roles.

So the number that decides everything is the ratio of salary dollars to AI dollars. At 50% of salary you get to $200B easily. At 10% you struggle to get there at all.

Jason Lemkin (@jasonlk), the founder of SaaStr, thinks it lands at $100K of tokens per engineer and dev teams 30-40% smaller. Run that across every software role and the US lands near $200B. Worldwide struggles to clear $350B.

Key takeaways already in your email via @PodWireHQ

Source: 20VC with @HarryStebbings

[PodWire @PodWireHQ]

@maxhodak_ Full episode: https://t.co/1qTCltzOeb

[PodWire @PodWireHQ]

Put electrodes in a paralyzed patient's motor cortex and they are using a computer within an hour. Drug discovery takes a decade to turn over one card.

Max Hodak (@maxhodak_), the founder and CEO of Science and formerly of Neuralink, says the difference is that the brain is a computer and biology is not. Making a drug means understanding molecular detail, and humanity just isn't that good at that.

His example of the hard road is CAR-T. Highly engineered, patient specific, and it can still hand you a giant immune overreaction.

Devices compound instead. Science's retinal prosthesis restored form vision nobody had produced in a blind patient before, and it is still black and white with a field of view like looking through a straw. Grayscale depth, then red and green, is an engineering roadmap rather than another coin flip.

Key takeaways already in your email via @PodWireHQ

Source: No Priors with @saranormous

[PodWire @PodWireHQ]

@AGraylin Full episode: https://t.co/AEOXqO14SC

[PodWire @PodWireHQ]

China delivers electricity at 2 to 3 cents per kilowatt hour, in some cases 10 to 15 times cheaper than many parts of the US, according to Alvin Wang Graylin (@AGraylin).

Graylin, a technology executive with 30+ years across the AI, XR, and semiconductor industries, says the grid buildout was never an AI project. It was about energy independence. 40% of China’s oil is imported, and he says the Hormuz disruption made the country glad roughly half its auto fleet had already gone electric.

The compute advantage falls out of that. Solar and wind go up in the western deserts, high voltage lines carry the power a thousand miles east and lose under 1%, and some data centers get built at the generation itself so no power strands.

He says China is now adding more new electricity generation than the rest of the world combined, about 10x what the US adds every year. He still thinks the constraint today is chips, not power. China cannot buy enough and cannot make enough without EUV machines.

Key takeaways already in your email via @PodWireHQ

Source: Moonshots with @PeterDiamandis

[PodWire @PodWireHQ]

@GrantLaFontaine Full episode: https://t.co/WFzXvpRkjQ

[PodWire @PodWireHQ]

Whatnot users spend 95 minutes a day on the platform, and more than 80% of them don't buy anything on a given day.

Grant LaFontaine (@GrantLaFontaine), co-founder of Whatnot, says most people aren't even purchasing. They're just there watching and having a good time.

a16z General Partner David George (@DavidGeorge83) calls 95 minutes a shocking number that puts Whatnot on par with social and entertainment platforms. His read is that with over 80% of people not transacting, it is clearly a form of entertainment.

LaFontaine's version of the buyer value prop is simple. Shopping on Whatnot is fun, and there are not a lot of fun experiences shopping online.

Key takeaways already in your email via @PodWireHQ

Source: The a16z Show with @DavidGeorge83

[PodWire @PodWireHQ]

@benthompson Full episode: https://t.co/lBC0wLDJU3

[PodWire @PodWireHQ]

Berkshire's BNSF railway threw off more free cash in one year than See's Candies had in its entire lifetime, and Ben Thompson (@benthompson) thinks Google $GOOG is making that same swap with AI.

Thompson, who writes Stratechery, says the trouble with a tremendously high margin business is not the margin. It is that there is nowhere left to put the money. The percentage profit is huge, the reinvestment runway is tiny, and the cash just piles up.

Buying a railroad fixed that. Way worse margins, but absolute dollar amounts so large that the worse margins throw off much bigger absolute profits.

Search is Google's See's Candies. Zero marginal cost, scales in every direction. AI is the railroad. It incinerates cash and its addressable market is basically all white collar work.

Which is why the equity issuance reads differently. A smaller percentage of an astronomically larger pie is not something shareholders end up complaining about.

Key takeaways already in your email via @PodWireHQ

Source: Invest Like the Best with @patrick_oshag

[PodWire @PodWireHQ]

@RichardSSutton Full episode: https://t.co/D0koBIqoOW

[PodWire @PodWireHQ]

Rich Sutton (@RichardSSutton) says LLM assistants do not actually learn. Their weights never change.

Asked whether today's assistants are continual learners, the man who pioneered reinforcement learning and wrote "The Bitter Lesson" answered "Are you serious?"

They keep memories about you. They do in-context learning. The weights never move.

All the structuring and generation of new concepts that went into building those models was the weight learning, and it ran exactly once. Sutton's objection is that you don't want it to happen just once.

Key takeaways already in your email via @PodWireHQ

Source: Training Data with @sonyatweetybird and @Alfred_Lin

[PodWire @PodWireHQ]

@glangley Full episode: https://t.co/fqdFs8jhYQ

[PodWire @PodWireHQ]

Flock Safety CEO Garrett Langley (@glangley) says a Georgia police chief fired 9 officers in one day for abusing Flock's license plate system.

The tell he describes is a cop who searches the same license plate day after day and never puts it on the hot list.

A real investigation does the opposite. You add the tag so the whole department gets notified, because you might be asleep or off shift when the car turns up. Someone stalking a person wants nobody else to know.

Langley says Flock shipped the tool 4 months ago and it has caught more bad cops than he expected. The check is now mandatory for customers rather than a default they can switch off.

Key takeaways already in your email via @PodWireHQ

Source: All-In with @Jason