On TBPN, the hosts discussed figures presented at Anthropic’s Investor Day and the effect of its efforts to combat unauthorized model distillation. One host said the company’s Q3 revenue was $16.8 billion and annualized the quarterly figure to about $67 billion. In another passage, he gave the figure as $67.2 billion. That is roughly ten times the level a year earlier, but, in his assessment, the market had expected a stronger result and may have been looking for $90 billion.

The hosts said Anthropic estimates that distillation accounts for 5–7% of its revenue. The host stressed that this was the company’s estimate, not an independently verified measure of losses. He described the potential impact as worth billions and suggested that combating distillation could weigh on near-term revenue growth while limiting the emergence of competing models over the longer term. According to the host, Anthropic said it had deployed its own frontier technology and believed it could stop distillation.

The discussion raised the possibility that operators could disguise large-scale querying as ordinary use—for example, by setting up hundreds of Max accounts, each with its own VPN, or by using third-party SaaS wrappers that route requests to model APIs. The hosts said such arrangements could make detection difficult: a model provider might see many requests coming from a seemingly legitimate company, while the wrapper itself might not be able to inspect document contents. These examples and their scale were presented as the hosts’ theories, not as verified facts about specific companies.

The hosts also considered whether Meta would be willing and able to fight distillation. One host speculated that Meta might have advantages in bot detection, but also said the company might be less concerned about strong open models because of its distribution channels. A participant said Meta had released the open model MuseSpark Open that day. In the participant’s view, this suggested Meta might have a different attitude toward the risk of an open model emerging through distillation. A host also argued that a model not attracting distillation attempts might not be at the frontier; this was his opinion, not an established definition.

Amid a decline in technology stocks, TBPN’s hosts sarcastically listed drastic measures: sell stocks, bonds and dollars; buy land, guns, food and medicine. One host predicted a near-term shift to a barter economy. In context, this was a hyperbolic market comment, not a specific investment recommendation.

The hosts linked the market mood to lofty expectations for exponential growth in AI-lab revenue. One host said expectations that Anthropic would grow revenue tenfold every year might not be met: in his estimate, revenue had risen from a little over $100 million to $1 billion and then to $9 billion, indicating some deceleration. He added that those expectations logically implied $100 billion in revenue, but he thought a figure around $80 billion was more likely.

The hosts also paraphrased a post by South Korean President Lee Jae Myung. According to one host, a translation from Korean produced by Grok read “we must live” and included a suggestion to call 109 if someone was struggling. The host said he did not know who the “Gil” mentioned in the translation was. He also said that, as far as he knew, the KOSPI was at a six-month low and described the situation as bleak. The translation and index assessment are reported as the participants presented them on air.

The hosts discussed expectations for Anthropic’s revenue and how difficult it is to sustain exponential growth. The public discussion cited a progression of roughly $100 million, $1 billion and $10 billion in revenue. One host clarified that the next reported step was from $1 billion to $9 billion—a 9x increase, not 10x—and said that reaching $100 billion from $9 billion would represent acceleration, not deceleration. At the same time, expectations of $100 billion or even $150 billion raise a bar that is difficult for a company to clear.

One host speculated that revenue could finish around $80 billion if the slowdown from $1 billion to $9 billion were extrapolated. He also argued that competitors and major technology companies in frontier AI could form alliances or make deals with other labs to prevent one leader from achieving a monopoly. This was the host’s analysis and speculation, not a confirmed plan by Anthropic or other companies. He noted that Investor Day is not the IPO itself but could be a precursor; in his view, labs may disclose less exciting figures ahead of a period of accelerating growth and a public listing.

The hosts recounted a statement by Amit on CNBC: the fund had sold all of its CoreWeave shares and held none at the time of the interview. He said the fund had reduced its exposure over the previous three months because it believed the AI market had overshot and was heading into a period of consolidation. The fund also reportedly moved 10% of its book into Meta and Tesla.

The hosts said declines in some AI stocks could reflect a rotation of capital rather than money leaving the market: they cited a Nasdaq decline of roughly 0.3–0.4% while the Dow was up almost 1%; in another comparison, the Nasdaq was nearly flat over five days. They suggested investors selling CoreWeave might be buying SaaS companies and other businesses with advertising revenue or strong margins. That was the hosts’ commentary on market dynamics, not evidence of specific purchases by investors generally.

One host said a rumor was circulating, attributed to Martin Shkreli, that the Vara AI fund (Value Aligned Research Advisors LLC) had blown up. The discussion did not confirm the claim; the host said they would follow developments.

For context, another host listed holdings shown in the latest 13F: Nebius, Bloom Energy, CoreWeave, Seagate, Western Digital, Hut 8, Iridium, Core Scientific, Nvidia and SanDisk. He characterized the portfolio as being further out on the risk curve. The hosts emphasized uncertainty about the fund’s condition and did not state that it had in fact collapsed.

The host recounted Blake Byers’s forecast about the effects of AI capital expenditure on markets. Byers called rising interest rates his modal prediction for the next decade, as capital shifts from savings and Treasury securities into investments promising higher returns from AI. In his scenario, this could cause short-term panic as higher rates increase debt-servicing costs for governments and other borrowers.

The host connected the hypothesis to developers that financed large projects at floating rates: some now have to refinance, sell property or halt construction. In Byers’s scenario, after three years of panic, investments begin to pay off substantially; economic revenue growth offsets borrowing costs, and panic turns to euphoria. He also predicted that developed economies could come to regard 5% growth as normal and then demand 10%, while healthcare and R&D spending rise as a share of GDP. This is a forecast, not an established outcome.

Nick Cammarata of Goodfire argues that modern tools are changing how researchers view neural-network interpretability. He says an early automated interpretability project at OpenAI made only about 2% of neurons legible with the primitive tools then available. Now, he says, the vast majority of neurons in models with up to trillions of parameters are interpretable, and many supposed examples of tangled representations were misunderstood. Cammarata predicts ambitious interpretability will be achieved by the time it is most needed, possibly quite soon as automated researchers improve.

The discussion described mechanistic interpretability as examining what happens inside a model at the level of individual neurons. The earlier superposition hypothesis held that a neuron might respond to many features; sparse autoencoders were used to recover a representation in which each feature has its own neuron. As a participant summarized the new work, that approach may no longer be necessary: researchers can inspect the neurons themselves. This could help not only with safety but also with improving model capabilities, such as mathematics.

The hosts called the work a potentially important scientific moment and suggested interpretability could make alignment more like a conventional science. The discussion did not independently verify Cammarata’s claims; they were presented as his conclusions and the hosts’ assessment. Goodfire was described as an independent mechanistic-interpretability lab. The hosts speculated it might work on open models or serve as an outside auditor for labs, but these were possibilities, not announced arrangements.

Ahead of an expected Tesla Roadster presentation that day, a host said he had placed a deposit and was ready to walk away if the car did not impress him. A participant gave the presentation time as 7:55 p.m. Central, or 5:55 p.m. Pacific; the time was not independently confirmed in the discussion. The hosts discussed a rumor that 7:55 might refer to a quarter-mile time, but it remained speculation. The host said he had paid $5,000 by credit card and would need to wire another $45,000 within 10 days. He said he could assess the presentation that night and keep his place in line if the Roadster looked compelling, or cancel and leave the line if it did not. He estimated that tying up $5,000 for two days at a 6% interest rate would cost about $5 in foregone returns.

Discussion of the car’s capabilities was speculative, not a report of confirmed features. One host predicted the demonstration might involve ramps and a jump; another bet that the car would rise more than 3 feet from flat ground, possibly more than 6 feet. The hosts also joked about rocket engines, thrusters or fans lifting the car, and a looped ramp sending it through the air before landing. The excerpts did not establish any actual Roadster capabilities.

The podcast hosts speculated about how Tesla might demonstrate the new Roadster: one possibility was a ramp jump, another was using rocket engines or thrust for a brief hover. One speaker expected a combination of flight and Nitro Circus-style stunts; another bet there would be no ramps. These were predictions, not claims about the car’s confirmed capabilities.

The discussion also revisited past delays: according to a host, buyers of the original Roadster paid in full in 2017 and waited about a decade without receiving the car. The speakers suggested that a failure to deliver the new vehicle within a year of its reveal would count as being “fooled again.” The $5,000 and $45,000 deposits mentioned were part of a hypothetical scenario, not confirmed order terms. One speaker also called Mission S the safer bet and said he thought Tesla would deliver it.

After a video of a car the speakers called the Porsche Mission S appeared online, the hosts discussed its design. One said it looked much more aggressive than expected and predicted it could put pressure on McLaren, Lamborghini and Ferrari in the mid-range supercar segment. Another pointed to the roof scoop and said the car evoked the Porsche Carrera GT and 918 Spyder, while noting that its performance remained to be seen.

One speaker called it the most desirable car in the $500,000-to-$2 million range—his personal judgment, not a measure of demand. The conversation also compared its design to the Porsche 959 and GT1, and described it as feeling much farther removed from the 911 than the Cayman does. The hosts speculated that the wide rear tires could throw stones and joked that drivers might need mud flaps. The actual reveal was expected in less than 90 minutes at the time; the discussion did not provide confirmed pricing or specifications.

Semafor’s San Francisco bureau chief, Reid, said he drives a Tesla Model Y and mainly wants more room for his children and gear. He was interested in the longer-wheelbase Model YL and said he planned to look into it when his lease ends. A host said he had ordered one, but delivery, expected in November, had been pushed to January because Tesla could not make them fast enough.

Another speaker suggested Tesla was reaching an inflection point as more people experienced what he described as the magic of FSD. He also speculated that Elon Musk was not giving Tesla enough attention, perhaps because he wanted to merge it with SpaceX and owns a larger share of SpaceX. That was the speaker’s interpretation, not a confirmed merger plan. The hosts also shared personal experiences: the guest activated FSD while his shoulder was injured and found it helpful to have the car drive him; he said FSD had helped him during a hockey injury for the second time.

The speakers observed that some people were buying Teslas despite political objections to Musk. One described FSD as seeming to some buyers like the safest choice for their families and others on the road. These were personal impressions and opinions, not sales data or independent evidence of safety.

A guest said he had recorded an interview with Microsoft CEO Satya Nadella for a new podcast scheduled to air the following week. One question was whether Microsoft could remain a great business in a few years yet fall out of the “Magnificent Seven.” As context, he cited Nvidia earning more per quarter than Microsoft and mentioned growth among hyperscalers and the arrival of SpaceX AI. Nadella’s answer was not included in this excerpt; the guest said only that Microsoft communications staff reacted with a small gasp.

Another speaker called Nadella, in his opinion, the most candid communicator among Magnificent Seven leaders, saying he seemed to answer questions directly rather than continually returning to set talking points. The guest also recounted Nadella’s remarks to security executives: according to the guest, Nadella expressed disappointment with models and doubts that alignment was working or that mechanistic interpretability would be achieved. As the guest relayed it, Nadella instead discussed putting guardrails and secure, deterministic software around models. This was a secondhand account, not an independent assessment of the models.

Discussing Xbox Labs and games assembled or decompiled with AI, the Semafor guest called Xbox head Aisha the right leader for the moment, describing her as more forward-thinking and flexible. He suggested the industry would have to rethink intellectual property if users could decompile and mash up games at home with tools such as Claude. He predicted this could accelerate cloud gaming in new ways and added that Gen Z likes it.

The host said such mashup games seemed fun as memes but not necessarily like future blockbusters. He also argued that being owned by Microsoft gives Xbox financial flexibility: unlike independent game companies, which depend on sales and quarterly results, Xbox can experiment with its business model. The guest agreed, while stressing that the console business remains significant and that Xbox would probably need to disrupt itself somewhat. These were the speakers’ assessments and predictions, not an announcement of a specific Xbox policy change.

A host noted that Mac minis, which were not originally built as AI devices, had been flying off the shelves as people began using them for AI tasks. He compared this with Nvidia graphics cards: GPUs found uses in AI, and the company later developed server chips.

The conversation also covered Xbox and other older hardware that users can take apart, modify, or use to run agents and AI-assisted coding tools. In the host’s view, a reliable, affordable product can find buyers even without a large intellectual-property margin. The guest added that he was searching through his drawers for old hard drives because he did not want to buy new ones. These were the speakers’ observations and opinions, not quantified sales data.

One speaker suggested that the shortage of consumer AI products may be due to limited computing capacity: there are not enough data centers to run frontier models at scale for consumer applications. In his view, open weights could enable the next stage before cloud-based AI services do: early adopters might run models on their own computers before a broader Web 2.0-like phase emerges.

In the same discussion, the speakers identified financing as another possible bottleneck. A participant who had discussed the issue with his finance team argued that technology companies’ borrowing was not a replay of the 2008 financial crisis or the 2000 bubble. He described the borrowing as opportunistic, but said banks could reach their lending limits; if so, money could constrain companies that need capital. In his assessment, the largest technology companies still have cash and other ways to fund investment. He said he intended to write about the issue in his newsletter and called the situation a “bottleneck economy.”

These were the speakers’ views, not confirmed evidence of an actual capital shortage or a prediction that financing has already run out.

Brian Cantrill, co-founder and CTO of Oxide Computer Company, said the company develops rack-scale computers whose hardware and software are designed together. Oxide was founded in 2019. Cantrill said the company was built on the belief that cloud computing should be available not only in public clouds but also in companies’ own data centers.

Cantrill connected that idea to his experience at Joyent. There, the team built a cloud on standard Dell, HP and Supermicro hardware; after Samsung acquired Joyent in 2016, rapid growth exposed problems at the hardware-software interface. Cantrill said major cloud companies such as Google, Amazon and Facebook had built their own rack-scale systems with co-designed components, while other businesses could not buy comparable infrastructure for their data centers.

According to Cantrill, Oxide raised $20 million in its seed round—an amount needed to hire a team and tackle a technically challenging project. He credited Eclipse and investor Seth as early believers in hardware-software co-design. He listed a custom compute sled, a switch and software spanning low-level firmware, the control plane and distributed storage systems among the company’s work. He also said the company completed its first instruction on x86 and eliminated dependence on UEFI. Cantrill described Oxide as “nine startups in one” and said the technical risk was high.

The founder of Advanced Spade Company said he started the company in 2024 after working at SpaceX and The Boring Company, on construction projects in Austin, and for about two years at Anduril. His interest in underground construction grew from the difficulty of knowing where existing utilities are and avoiding damage to them. After wildfires in Los Angeles, he said, people began coming to the team asking how power lines could be moved underground. The company digs holes and installs utilities.

Advanced Spade began with private utility locating and mapmaking. General contractors, which the founder said typically hold contracts from utilities, later asked the company to do drilling work. He said Advanced Spade uses its own datasets and maps to gain an edge when drilling, and expands the size and technical difficulty of its projects gradually. In his account, growth in the industry depends on hands-on fieldwork and reputation; technology helps the company move faster through the construction industry’s usual growth curve.

The company combines subcontracting with a gradual expansion of its scope. The founder said subcontracting involves less administrative burden but gives the company less control over its own direction. As it moves into more difficult work, Advanced Spade wants to leave lower-end work and sell other, smaller contractors tools developed through its own projects. He cited a locator that the team has, he said, spent years improving. In his view, such a tool is easier to sell after the company has used it while competing for contracts.

The founder also said the company has engineers integrating data and building better tools. Asked about new technologies and materials, he said there are ideas he looks forward to discussing in about five years. This is an expectation of a future disclosure, not evidence that those developments are already ready.

In an interview, a company representative whose name is not given in this excerpt described its approach as a combination of integrating existing technologies and building some solutions in-house. Customers sometimes characterize the company as a “slick integrator,” he said, but it has no SaaS costs or extra software layers to pay for and can build the tools it needs. Some challenges are mostly integration, he added, while developing a new locator is difficult.

The company configures purchased equipment for its needs; one of the machines shown was a directional drill. The guest linked demand for underground infrastructure to problems with overhead networks: in his account, investment in underground grids has tripled over the past 10 years, while overhead infrastructure continues to fail. Trenchless directional drilling became widespread in part because it can route networks beneath roads without digging them up. But he said the available tools and technology are not yet sufficient for safe work in cities, where crews risk hitting existing underground lines. He also associated the field’s progress with the data-center boom and lower costs. The company has roughly 50–55 employees, he said, and confirmed that it is hiring.

The guest said there are new technologies requiring advances in materials that he hopes to discuss “in like five years.” That was a promise to discuss them in the future, not a claim that the capabilities already exist.

The host discussed a clip of the racing simulator Assetto Corsa with mods, including Nvidia’s DLSS 5, which he described as AI upscaling. He said he could not confidently distinguish the image from real video and called it a video that had “passed the Turing test.” He also qualified that assessment: despite trying to verify the clip, he could not say with complete certainty. The other speaker said graphics were not especially important to him but that he would try the feature. These were the hosts’ impressions of the clip and the effect’s potential appeal, not an announcement that the feature had launched in a commercial game.

The host said Reese Witherspoon was buying her company, Hello Sunshine, back. He recalled that the company was valued at $900 million in 2021, but described the possible current price inconsistently: as “94%” of that valuation and then as roughly $7 million. Those figures do not align, so the excerpt does not establish the deal price. The host attributed Candle Media’s difficulties to slower-than-expected growth in direct-to-consumer subscriptions and said he was rooting for Witherspoon to regain full ownership. The transaction and its terms were described as possible, not confirmed as final.

Discussing startups’ paid partnerships with popular anonymous accounts, the hosts said such advertising could put audiences off and give a brand a reputation for artificial promotion. They noted that several companies apparently offer services arranging these placements. One host, however, praised the disclosure of paid partnerships on X: he said paid posts had previously appeared on the platform without disclosure and called the labeling a step toward greater transparency. These are the hosts’ comments, not confirmation of a specific platform-policy change in the excerpt.