Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter

Note: This is a summary of the speakers’ arguments, claims, predictions, and opinions. It does not independently verify them. The transcript contains repeated caption lines, misspelled names, and some unclear figures.

• All-In Podcast, episode 278.
• David Friedberg is absent.
• Guests are Travis Kalanick and investor Gavin Baker.
• Gavin briefly discusses the excitement around the SpaceX IPO and adopting a “0–0” mindset: ignore past victories and treat the next challenge as a fresh start.

The hosts describe a “socialist sweep” in three New York Democratic congressional primaries involving candidates endorsed by Mayor Mamdani:

• Brad Lander defeated incumbent Dan Goldman in New York’s 10th District.
• A 32-year-old democratic socialist candidate defeated a long-serving incumbent backed by House Democratic leadership in the 13th District.
• Claire Valdez won the open 7th District seat over the incumbent’s chosen successor.
• The districts range from wealthy parts of Manhattan and Brooklyn to poorer areas of Harlem and the Bronx.
• Because these are safe Democratic seats, the primary winners are expected to win the general election.
• The speakers say the candidates performed especially well among younger, college-educated and relatively affluent voters.
• They mock this group as people wealthy enough to support socialism while presenting themselves as economically disadvantaged.

Chamath argues:

• Silicon Valley has badly presented capitalism and AI to ordinary people.
• Industry leaders allowed their rivalries, insecurities and personal disputes to become public, damaging their credibility.
• This gave socialists an opening to portray almost any alternative as better than the present system.
• Some recent congressional races were effectively referendums on AI, with anti-AI and pro-AI political groups financing opposing candidates.
• Pro-AI candidates narrowly held some important seats.

Chamath describes AI as potentially the greatest economic equalizer of their lifetimes:

• Google and internet search made the world’s information accessible.
• AI goes further by converting information into usable expertise, intelligence and action.
• It can give every person something resembling a highly capable co-founder, engineer, strategist or problem solver.
• Access is less dependent on conventional gatekeepers.
• A person’s results will increasingly depend on how well they direct AI toward productive goals.
• He dismisses much of the discussion about job destruction and water consumption as misinformation promoted by interested parties.
• His larger complaint is that the technology industry has failed to explain AI’s benefits compellingly.

Sacks presents the DSA as seeking a fundamental reconstruction of the American system. He says its programme includes:

• Abolishing the Senate.
• Abolishing what it calls the “carceral state,” including prisons and conventional policing.
• Abolishing ICE, granting broad amnesty and opposing deportations.
• Making the executive and judiciary subordinate to Congress.
• Ending the Electoral College.
• Expanding the House of Representatives.
• Introducing proportional representation, ranked-choice voting and a multiparty system.
• Public ownership of major corporations.
• Supporting Palestine.
• Cutting or eliminating the military department.

He characterizes these ideas as revolutionary rather than ordinary progressive reforms.

The speakers repeat a series of allegations concerning one newly nominated candidate, including alleged statements against Western civilization, the American flag, police, prisons, borders, military personnel and private ownership.

They use these alleged comments to argue that:

• Some incoming DSA members are not merely critical of America but reject its constitutional and economic system.
• They are considerably more radical than older progressive figures.
• Democratic congressional leaders may struggle to control them.

These are presented in the transcript as the speakers’ allegations, not independently established facts.

The panel quotes DSA figures as saying that they use the Democratic Party primarily as:

• A ballot-access vehicle.
• A convenient label under which to contest elections.
• A caucus they will cooperate with when useful.
• An institution to pressure or overcome rather than their true political home.

Their conclusion is that the DSA is attempting to capture the Democratic Party from within, much as Trump captured the Republican Party.

Sacks claims the DSA coalition combines:

• Recent immigrants.
• Young, highly educated progressive organizers.
• Downwardly mobile graduates working in universities, nonprofits and advocacy organizations.

He argues that Democratic support for mass immigration, multiculturalism and reduced emphasis on assimilation unintentionally strengthened this coalition.

He further argues that importing large populations without integrating them into American civic culture could gradually change the nature of the constitutional and economic system.

The panel briefly contrasts this with Trump’s proposals to eliminate agencies, saying Trump sought to remove particular agencies rather than abolish the Constitution itself.

As an aside, the speakers claim:

• USAID operated as a mechanism supporting left-wing organizations and political activity.
• Recent rightward electoral changes in South America may partly reflect the reduction of that support.

Again, this is presented as the panel’s interpretation.

Travis offers two broad ideas.

Truth and justice are society’s immune system:

• When truth becomes difficult to establish, social problems spread.
• When crimes have no consequences, social trust deteriorates.
• Watching whether truth and justice are becoming stronger or weaker can indicate society’s direction.

The desire for communism exists within ordinary human behaviour:

• Most people sometimes feel lazy.
• Most people sometimes want something without earning it.
• The danger arises when a system encourages those impulses without consequences.
• Once enough people participate, the system can become self-reinforcing.

Gavin argues that the historical Democratic Party attempted to represent:

• Working-class people.
• Poorer Americans.
• Black Americans.
• Hispanic Americans.
• People seeking greater economic opportunity.

He says the modern DSA is instead concentrated among relatively affluent, educated white liberals whose economic prospects are declining.

He believes the DSA is losing working-class and minority support while attracting graduates who expected elite careers but entered nonprofit, academic or advocacy work instead.

Gavin alleges that:

• Governments increasingly direct public money through nonprofits.
• These organizations can reward political allies with highly paid jobs.
• Spending can rise while the stated problem worsens.
• Homelessness spending in California and New York is his main example.
• This creates something resembling organized patronage or corruption.

He compares it to the “Curley effect”: politicians adopt policies that drive opponents or productive residents away, leaving behind a more politically dependable electorate.

Gavin says Mamdani’s success is primarily caused by political talent rather than socialist policy:

• He is charismatic.
• He speaks well in interviews and public events.
• He understands social media.
• He can adapt his presentation to different audiences.
• Gavin considers him substantially more skilled than AOC as a communicator.

The speakers believe his message resonates because younger generations feel that:

• Housing is unattainable.
• College debt is excessive.
• Healthcare is unaffordable.
• Inflation has damaged their prospects.
• They may perform worse economically than their parents.
• The existing economic system feels closed or rigged.

They argue that “democratic socialism” presents communism through a softer and more attractive political wrapper.

Chamath praises moves in Canada, Britain and Australia to restrict social-media access for people under 16.

He argues that:

• Social media radicalizes children before they can properly evaluate political information.
• Delaying access may produce healthier and less extreme future voters.
• Restrictions could eventually improve political stability.
• Similar policies should be considered across the United States.

Travis and Gavin agree that excessive social-media use harms children but warn that age verification can become a route to:

• Requiring adults to identify themselves online.
• Eliminating anonymous speech.
• Building national identity systems.
• Censoring political disagreement.
• Allowing governments to define criticism as “harmful content.”

They support protecting children but say anonymity and free speech must also be preserved.

Gavin argues that progressive intentions often produce measurable harm. He cites claims involving:

• Crime and mortality under different district attorneys.
• A small group of repeat offenders committing a disproportionate amount of violent crime.
• Environmental regulations preventing solar projects from being built.
• Eliminating selective schools and weakening mathematics education.
• Educational changes disproportionately harming low-income students.

His central point is that free speech remains essential because it allows the public to compare political intentions with measurable results.

The panel says the DSA is especially effective in:

• Low-turnout primaries.
• Local elections.
• Ranked-choice systems.
• Organized ballot collection and voter mobilisation.

They point to Los Angeles as a possible next test and say DSA-aligned organizers have gained influence on the city council.

They predict that even Democrats who are not DSA members will move left because they fear being challenged in future primaries.

The hosts interpret Brad Lander’s victory over Dan Goldman as substantially driven by Israel and Gaza:

• Goldman strongly defended Israel.
• Lander publicly described events in Gaza as genocide.
• Both candidates otherwise had similar progressive credentials.
• The result is presented as evidence that Israel has become a decisive issue in Democratic primaries.

The transcript claims that a large majority of Democrats now disapprove of Israel’s conduct.

The speakers say:

• Younger Democrats strongly oppose Israeli policy.
• Younger Republicans are also more critical than older Republican voters.
• Older Republicans and traditional cable-news audiences remain more supportive.
• The issue is creating arguments within both parties and among conservative media personalities.

Chamath emphasizes that criticism should distinguish among:

• Jewish people.
• Israelis.
• The state of Israel.
• Benjamin Netanyahu and his government.

He says it is reasonable to criticize Netanyahu without condemning Israel or Jewish people collectively.

He suggests that Israel may need new leadership to repair its international standing.

Chamath compares the political consequences of October 7 with those following September 11:

• A major attack creates emotional and psychological injury.
• Citizens temporarily give leaders unusually broad authority.
• Governments then take actions that would normally be politically impossible.
• The Patriot Act is given as the American example.
• He believes that the initial need to respond can later develop into excessive or prolonged action.

Gavin argues that Israel lacks effective international communicators:

• It needs young, fluent, social-media-capable spokespeople.
• It needs different spokespeople for the United States, France, Germany and other countries.
• Military heroes and senior officials may be admirable but are not necessarily effective online communicators.
• Israel is receiving continuous public-relations attacks without answering them effectively.

Gavin also argues that:

• American education often teaches children to feel ashamed of their country.
• America’s failures should be taught honestly, but alongside its accomplishments.
• Slavery and injustice should be placed in global historical context.
• America remains an unusually successful multicultural society.
• Immigration should preserve personal cultures while also producing participation in a common American culture.

The panel sees Mamdani’s strategy as similar to Trump’s:

• Build a movement outside the party establishment.
• Use the party’s ballot infrastructure.
• Communicate directly and emotionally.
• Expand the political tent without insisting on complete ideological agreement.
• Force established politicians to accommodate the movement.

The hosts concede that Mamdani is an exceptionally effective speaker. One says that a speech he gave about the New York Knicks was so compelling that it briefly overcame the host’s strong political dislike of him.

Their conclusion is that political communication and charisma can outweigh policy details.

The second major topic is Z.AI’s Chinese open-weight model, GLM 5.2.

The transcript describes it as:

• A frontier-class open model.
• Approximately 744 billion parameters.
• A one-million-token context window.
• Released under the permissive MIT licence.
• Downloadable, forkable and self-hostable.
• Available without regional restrictions or dependence on an API.
• Usable as the basis for commercial products.

The speakers say it scored highly on an intelligence index, surpassed GPT-5.5 on one software-engineering benchmark, came within roughly one percentage point of Claude Opus 4.8 on another measure, and costs substantially less to operate.

Gavin believes Chinese developers have extensively distilled American models.

He explains the alleged method:

• Operate very large numbers of accounts and devices.
• Send carefully designed questions to frontier-model APIs.
• Collect the complete reasoning traces or generated tokens.
• Use those answers as training data.
• Reproduce much of the frontier model’s behaviour at a fraction of its original development cost.

The panel compares this to copying the results of every conceivable Google search and using them to construct a competing search engine.

Gavin says GLM may now be capable enough to conduct much of its own reinforcement learning, reducing its dependence on further copying.

The panel acknowledges uncertainty:

• Upcoming American models have not yet been released.
• The performance gap could widen again.
• Nevertheless, the currently available Chinese model is already close enough to change competitive assumptions.
• A Chinese developer predicts open models will reach the next American frontier level by early 2027.

Gavin predicts that companies will not rely on one model. Instead, they will use a “council of LLMs”:

• A router examines each task.
• A company’s own open-weight model handles routine work and proprietary data.
• Frontier models handle the hardest questions.
• Frontier systems may also review or improve the open model’s answer.
• Several models cooperate as parts of one workflow.

He calls this a composable-model future.

Most requests may eventually run through cheap internal models, while only a small percentage use expensive frontier systems.

Gavin argues that open models:

• Reduce the margins available to frontier laboratories.
• Increase demand for chips, data centres, hosting and infrastructure.
• Are therefore beneficial to infrastructure providers.
• Do not eliminate the role of frontier laboratories, which may remain responsible for the hardest reasoning and verification.

The speakers suggest Nvidia could release a leading American open model whenever it chooses.

Reasons it may hesitate include:

• Competing with its own customers.
• Creating channel conflict.
• Damaging relationships with frontier laboratories that buy Nvidia chips.

They speculate that if OpenAI increasingly develops custom chips with Broadcom, Nvidia may feel freer to compete directly in the model business.

The transcript says Anthropic’s next model and OpenAI’s next model have been delayed or subjected to additional approval because of safety and jailbreak concerns.

Sacks argues that:

• Anthropic’s own public warnings encouraged officials to react aggressively.
• Anthropic has advocated an FAA-like federal regulator for AI.
• The resulting approval system could give established companies a regulatory moat.
• Anthropic may have obtained part of what it requested, although it is probably unhappy that its own model was delayed.

The speakers oppose creating an elaborate model-approval bureaucracy because it could:

• Reward large incumbents.
• Make entry difficult for competitors.
• Slow American releases.
• Give China additional time.
• Transform safety policy into regulatory capture.

They favour returning models to the market after specific vulnerabilities are corrected.

Sacks accepts that advanced models create cybersecurity risks but rejects permanent suppression.

His proposed response is:

• Give capable models to trusted cybersecurity professionals.
• Let “white hats” identify vulnerabilities before hostile users do.
• Patch vulnerable software rapidly.
• Use AI as a defensive force multiplier.
• Complete a broad security-upgrade cycle before China obtains equivalent capabilities.

His argument is that restricting American models will not restrict Chinese models.

The speakers estimate that China may remain:

• Several months behind in model quality.
• Much further behind in advanced silicon.
• Yet only a few months behind in practical overall capability.

They interpret this as evidence that American self-restraint could erase its hardware advantage.

The panel discusses claims that:

• Chinese laboratories are being pushed to train on Huawei chips.
• Recent Chinese models may have been trained using Huawei Ascend clusters.
• Their inference software is being optimized for Huawei hardware.
• China intends to reduce dependence on Nvidia and other foreign suppliers.

The speakers acknowledge that the claims might be exaggerated or conceal smuggled Nvidia hardware.

They predict China will combine:

• Huawei chips.
• Chinese open-weight models.
• Optimized inference software.
• A complete packaged system.

China could then export these “AI in a box” products globally at lower prices.

Sacks argues the United States should have exported its technology more aggressively to allies before China developed a competitive alternative.

The hosted Chinese model reportedly avoided answering questions about Taiwan and Tiananmen Square.

The panel says:

• Hosted Chinese services will retain state-directed censorship.
• Open weights allow American companies to fork the model.
• Developers can add omitted historical information and remove some restrictions.
• Therefore, censorship is a serious limitation but not necessarily fatal to the model’s international use.

The third major topic is Micron.

The transcript says:

• Micron, Samsung and SK Hynix are the three major manufacturers of high-bandwidth memory.
• Micron’s 2026 supply is sold out.
• Its reported revenue rose dramatically.
• Earnings and guidance exceeded expectations.
• Micron’s share price had risen many times over.
• Gavin had previously identified HBM manufacturers as a major investment opportunity.

Gavin argues that the meaningful bottleneck is not an obscure component but DRAM:

• AI performance depends heavily on memory capacity.
• It also depends on memory bandwidth.
• Every advanced accelerator requires large amounts of specialized memory.
• HBM is becoming increasingly customized.
• Advanced AI memory is extremely difficult to manufacture.
• Elon Musk’s proposed fabrication strategy is therefore focused heavily on memory.

HBM is made by vertically stacking multiple DRAM dies:

• Earlier generations used approximately eight layers.
• Newer systems may use 12 or 16.
• Aligning, connecting and packaging the dies is itself an advanced manufacturing challenge.
• Only a handful of companies can reliably produce it at scale.

Micron has introduced agreements containing price floors and ceilings with major customers.

Gavin says:

• A small number of large buyers may account for roughly half its revenue.
• Contract price floors are reportedly above the profitability peaks of earlier memory cycles.
• This could make memory manufacturing less cyclical and more structurally profitable.
• Memory companies may eventually receive valuations closer to other critical semiconductor suppliers.

China’s CXMT may expand the supply of cheaper consumer DRAM.

However:

• Consumer-grade DRAM is easier to manufacture.
• HBM and other AI-server memory remain much more difficult.
• CXMT may eventually reduce pressure on Apple and consumer electronics.
• It is less likely to solve the immediate advanced-AI memory shortage.

AI data centres are absorbing a growing share of global memory production.

The panel says this is increasing prices for:

• High-memory MacBooks.
• Mac Studios.
• Xbox consoles.
• PlayStation.
• Nintendo Switch.
• Other consumer devices.

Apple is described as passing more of the increased memory expense to buyers.

Gavin estimates that DRAM could become 30–40% of hyperscaler capital expenditure.

He says this may have a socially useful side effect:

• More expensive hardware makes unlimited data-centre construction harder.
• Companies must pay closer attention to economics.
• Society gains additional time to adjust its employment systems and social contract to AI.

The panel says:

• A conventional advanced fab may take two to three-and-a-half years.
• Environmental and permitting rules can delay construction even when a state offers large subsidies.
• Micron’s New York project is used as an example.
• States such as Texas may become more attractive because physical construction is easier.
• The decisive incentive is increasingly permission to build, not merely financial assistance.

They expect Elon Musk to attempt a faster process, potentially using an Intel partnership, while acknowledging that memory fabrication remains extraordinarily difficult.

The panel rejects the assumption that data centres will continuously become cheaper.

They identify increasing costs involving:

• Memory.
• GPUs.
• Transformers.
• Switchgear.
• Cooling equipment.
• Skilled labour.
• Power connections.
• Land-use approvals.
• Political opposition.

One estimate given is approximately:

• $35 billion in semiconductors for one gigawatt of compute.
• $25 billion for terrestrial power and cooling infrastructure.

Gavin argues that reusable Starship launches could eventually make orbital data centres economically competitive.

His model is:

• The silicon costs roughly the same in orbit as on Earth.
• Launching a gigawatt of equipment might eventually cost around $5 billion.
• The total could be roughly $40 billion in orbit versus $60–70 billion terrestrially.
• Terrestrial labour, power and construction costs are inflationary.
• Reusable-launch costs should become deflationary.

He says orbital data centres should be imagined as racks distributed through space and connected by lasers, not as one enormous building in orbit.

The speakers predict:

• More proposed data centres will face local opposition.
• Electricity demand will far exceed available supply.
• Large existing sites with dependable power will become extremely valuable.
• Sites containing hundreds of megawatts or more will resemble scarce “diamonds.”
• The shortage may accelerate SpaceX’s orbital-compute advantage.

The hosts again speculate that Tesla and SpaceX may eventually combine.

They treat the appearance of a recent Tesla “Megapod” trademark as possible evidence of a broader integrated energy-and-compute strategy.

The trademark describes a self-contained modular data-centre system containing:

• Computer servers.
• AI-processing hardware.
• Networking.
• Power distribution.
• Cooling.

The panel interprets this as a shipping-container-style data centre that can be:

• Built in a factory.
• Delivered by truck.
• Lowered onto a concrete pad.
• Connected to electricity.
• Activated within roughly 90 days.

One rumour is that Megapods could be installed near Tesla Supercharger locations because those sites already possess:

• Land.
• Electrical connections.
• Existing approvals.

Obstacles include:

• Physical security.
• Public access.
• Liquid cooling.
• Mechanical cooling.
• Reliability guarantees.
• Information-security requirements.

Travis says his business controls hundreds of properties with:

• Electricity.
• Mechanical cooling.
• Natural-gas connections.
• Commercial kitchens.

He is considering placing some internal robotics and AI compute in those locations.

The others distinguish between:

• Running hardware for one’s own company.
• Selling compute to outside customers.

External cloud customers would require service-level guarantees, security controls and potentially financial penalties for outages.

The panel says geographically distributed training is generally inefficient:

• Training requires rapid communication between GPUs.
• Even a few kilometres of separation can create unacceptable delays.
• Large training clusters need tightly connected hardware.

Distributed inference is more promising because workloads can often be separated and routed more flexibly.

The speakers divide inference into:

• Prefill: understanding the prompt and the conversation or document context.
• Decode: generating each successive output token.

Prefill benefits from greater memory capacity. Decode depends heavily on memory bandwidth and rapid token production.

Separating these workloads may allow different chips to handle each stage.

Older Nvidia A100 and H100 GPUs could remain useful by pairing them with specialized decode systems from companies such as Groq or Cerebras.

This could:

• Make older data centres competitive.
• Keep GPUs economically useful for seven to twelve years.
• Reduce depreciation.
• Lower financing costs for AI infrastructure.
• Support modular Megapod-style facilities.

The panel mentions projects including Bittensor, Venice, Plurales and Targon.

Their general model is:

• Individuals contribute unused computing capacity.
• Validators confirm that contributors provide the hardware they claim.
• Users rent compute on a permissionless market.
• Distributed infrastructure supports open models and reduces centralized control.

Chamath sees distributed inference as useful for:

• American competitiveness with China.
• A strong open-source ecosystem.
• Free speech.
• Reducing dependence on a few large cloud companies.

The panel imagines a future Tesla system in which:

• Every Powerwall includes AI chips.
• Customers receive a discount for accepting the hardware.
• Tesla uses or rents the chips when the homeowner does not need them.
• Tesla pays the owner for available battery and compute capacity.
• Starlink provides networking.
• Millions of homes become a distributed inference network.

They also briefly mention a rumour that Elon Musk could buy T-Mobile.

The final topic concerns public offerings by:

• SpaceX.
• Cerebras.
• OpenAI.
• Anthropic.

The hosts ask whether public markets can absorb several trillion dollars of combined valuations.

Gavin says only a small portion of each company will actually be offered, so global capital markets can absorb the available shares.

Gavin makes an exceptionally bullish prediction:

• Anthropic could be worth approximately $3 trillion as a public company.
• He expects annual revenue to exceed $100 billion.
• A later year could reach $200–300 billion.
• Inference could produce gross margins near 85%.
• At that scale, the business could be highly profitable.

This is explicitly Gavin’s valuation opinion, not a verified market price.

Gavin disputes the idea that SpaceX’s lockup expiration will automatically produce massive selling:

• Employees and investors have received recurring opportunities to sell private shares for years.
• Many therefore already achieved liquidity.
• A large portion of employees reportedly bought shares in the IPO.
• Many outside investors hold shares through special-purpose vehicles that may not distribute immediately.
• Elon Musk and several large holders are not expected to sell.

Others note that a very large one-year increase in valuation could still encourage some holders to take profits.

Cerebras fell below its deal price after reporting its first public quarter.

Gavin explains a market mechanism:

• Some portfolio managers automatically sell any IPO that falls below its issue price.
• They treat the broken deal price as a failed promise.
• Short sellers know this.
• They may push a stock just below the issue price to trigger automatic selling.
• That creates a self-reinforcing pile-on rather than a purely fundamental judgment.

Gavin advises companies to price IPOs so the stock is unlikely to break issue price during its first nine months.

The panel says:

• Excessively aggressive pricing can damage a good company.
• Trading below issue price does not necessarily mean the underlying business is bad.
• Facebook and Rivian are mentioned as previous examples.
• Public-market communication differs greatly from venture-capital fundraising.

Gavin says Cerebras should explain the delay between signing a major contract and recording revenue:

1. Order additional wafers from TSMC.
2. Wait roughly four months for chip production.
3. Spend another two months building servers.
4. Find electrical capacity.
5. Energize the hardware.
6. Begin producing tokens and revenue.

Therefore, a large OpenAI agreement signed in December would not materially affect reported revenue until the following third quarter or later.

For Cerebras, Gavin says the key question is not merely model competitiveness or customer demand.

It is:

• How quickly can the company bring new megawatts online?
• Can it obtain power?
• Can it deploy servers?
• Can it convert each megawatt into billable token production?

He estimates that adding around 50 megawatts per month during 2027 could create a cloud-computing revenue run rate near $9 billion.

He warns that only a small number of companies have successfully activated data-centre capacity at gigawatt scale.

Chamath strongly favours auctions:

• Investors state how much they will buy at different valuations.
• The company chooses the price that clears the desired amount.
• Investors receive another opportunity to improve their bids.
• The market, rather than bankers’ preferences, establishes the final valuation.

They recall using a Dutch auction for an early Uber financing:

• The round began around a $9–10 billion valuation.
• It ultimately cleared near $17 billion.
• Approximately $1.5 billion was raised.
• At the time, that was considered an extraordinary private-company valuation.

Gavin’s conclusion:

• A multitrillion-dollar valuation does not mean trillions of dollars must immediately be raised.
• A company may sell only a small percentage of its shares.
• A $50–150 billion offering is large but manageable in global public markets.
• Much of the value is already held in private markets; an IPO merely transfers it into a larger public market.
• Today’s capital numbers seem enormous compared with what founders considered extraordinary a decade earlier.

The episode closes with jokes about how difficult it once was to raise even a few million dollars, family stories about carrying hot potatoes to school for warmth and food, and the standard podcast outro.