A host described an informal poll of several iPhone Duo owners. Some said their total screen time might not have fallen, but their phone use had shifted away from vertical video. They said the compact outer screen worked well for messages and quick tasks, while opening the device provided an iPad-sized screen for Instagram; on the smaller display, Instagram felt less addictive.

The host stressed that these were only anecdotal observations and suggested checking them against Apple Screen Time data, comparing weekly usage before and after purchase and looking separately at Instagram and vertical video. He also said he saw an attendee at the event use two iPhone Duos rigged together to create content. He did not know why the attendee had set them up that way.

A host said Apple’s glowing logo on the MacBook Pro lid is back; the presentation’s opening light teaser pointed to that feature. The panelists noted that the light had previously disappeared, possibly to save money, and called the logo an iconic part of the Mac experience. According to the host, the logo’s return drew the loudest applause at the event—even more than the touchscreen.

The host estimated that the lighting costs about $2 per device. He stressed that this was a very rough estimate and reasoned that, amid rising memory and chip prices, that cost may now be less significant relative to the laptop’s overall price and component costs. One panelist suggested the light could be useful during livestreams by making the logo visible to viewers, but that remained speculation in the discussion.

One host speculated that lighting the Apple logo costs about $2 per device, stressing that this was a very rough estimate. He reasoned that if MacBook prices have risen from roughly $1,000 to $4,000 because memory and chips cost more, the $2 represents a smaller share of the retail price. In his view, the lighting technology probably has not become more expensive and may even have become cheaper. Other participants described the lighting as an unnecessary feature that adds visual appeal; the host saw it as part of Apple’s premium brand.

Google introduced Genie 4, describing it as its most realistic and interactive world model. According to the description, it can steer dynamic environments in real time—including events, camera angles, and actions—generate audio matched to a world, simulate photorealistic scenes and real-world physics, and create more coherent multi-step events through Gemini reasoning. Google has begun rolling it out as a research preview; one participant said they did not yet have access and suggested an application to an early-access program might be required. The hosts discussed possible uses in video games and creator tools, such as virtual backdrops and real-time set extensions. These were forecasts and examples of potential applications, not confirmation that commercial tools with these capabilities are already available.

Google introduced Gemini 4 Carbon as its next frontier model, following Argon’s release just two weeks earlier, according to the company statement recounted by the hosts. They said Carbon ranked first, with a score of 59, in the updated V5 Artificial Analysis Intelligence Index. They also cited cybersecurity evaluation scores of 88.2% on SCC Bench Pro for finding vulnerabilities and 54.8% on ExploitGym. One participant called the launch strong but said that for about a year, impressive benchmark results had repeatedly been followed by less impressive lived user experiences. Another praised Gemini 3 for concise, easy-to-scan answers. The hosts also speculated that Google might release an even larger model called Helium; this was conjecture, not an announced product.

The hosts said OpenAI plans to buy Google chips as part of its computing buildout. They also recounted a SemiAnalysis report that OpenAI’s Ohio lease commitment totals 8 GW across multiple phases, with Nvidia backstopping 4.25 GW. The report described it as the largest single leasing commitment it had ever seen. According to SemiAnalysis, OpenAI is also discussing GPU-as-a-Service and TPU-as-a-Service with cloud providers, is looking for sites, and may make a 1 GW TPU commitment. The hosts noted that some of these details were the source’s estimates rather than confirmed agreements. They framed the moves as an example of AI labs using different compute suppliers despite companies’ past disagreements.

According to information the hosts attributed to SemiAnalysis, Anthropic leased 1.7 GW of computing capacity this quarter, including 360 MW in Queensland, Australia. The agreement was signed on September 16, and operations at the site are expected to start in the first quarter of 2028. The source called it the largest overseas lease by an AI lab that quarter. Google backs 1 GW of the total; according to the source, it was not yet clear who would provide credit support for another large site.

One host argued that AI labs are using different chips, data centers, hyperscalers, and neoclouds for different models and stages of inference. In his view, even if one accelerator is faster and another handles more demanding tasks better, each can have an optimal use. He added that AI agents may be relatively easy to port across platforms, citing earlier coverage of an Anthropic implementation for AMD that, he said, was quickly built using vibe coding. The host viewed this trend as good news for investors expecting a broader set of silicon suppliers, but stressed that the market remains very early and heavily dependent on Nvidia.

The hosts quoted Dax of Open Code as saying that almost all DeepSeek traffic routed through OpenRouter the previous day could be served by one GB300 rack. He valued that volume at $6–7 million and said that even if hidden demand were ten times higher, compute costs would still be below $100 million. Participants questioned whether those figures captured the whole market: some users could access providers directly, and the estimate did not include companies’ own on-premises compute. Other possible serving channels named included Baseten, Modal, Fireworks, Together AI, Lambda, and many cloud platforms. One host suggested that these providers may account for the larger share of billions in inference spending, while noting that OpenRouter’s share of DeepSeek inference was impossible to know.

The hosts discussed an eBay listing for 300 racks of an AI supercomputer equipped with 9,600 NVIDIA A100 GPUs. The asking price was $60 million; participants estimated fair value at roughly $50–55 million. They said the hardware originally cost nearly $300 million and described the A100 as a chip about seven years old. One host jokingly suggested that anyone starting a neocloud provider could take out a $60 million credit line and buy the system. This was a joke, not a claim that a sale had occurred. Participants also said the item appeared in six users’ carts and that two units were available.

In a discussion of compute infrastructure, the hosts returned to Malaysia’s role as a regional data-center hub. One recalled an early Wall Street Journal report that described the country as a new data-center center and an “AI Switzerland”: in his account, Malaysia could buy compute in the United States and resell it in China, while government policy and available power supported infrastructure construction. He also suggested that proximity to Singapore enables a division of roles, with Singapore serving as a financial and port hub and equipment being installed in Malaysia. The host speculated that Chinese AI labs may use Malaysian neoclouds running Nvidia hardware because there are fewer trade restrictions there. This was the participant’s assessment, not verified information about particular customers or volumes.

Discussing Malaysia as an AI industrial hub near Singapore, a host speculated that Chinese AI labs training models may use Malaysian neocloud providers that rack Nvidia chips. He suggested that less restrictive trade controls could make this possible, but expressed uncertainty. This was the host’s speculation, not a confirmed report of specific providers or chip shipments.

Sequoia announced No Limit, a bot-building tournament hosted by Citadel, OpenAI and Sequoia. Teams of one or two will have 60 minutes to build a poker bot from scratch; agents will then play thousands of games across six scored rounds. The announcement described the contest as open to university students and recent graduates. The hosts noted that this eligibility rule excludes people who dropped out, joking that it was aimed at people who left academia to work on podcasts. One host wondered which tools and API the contestants could use, how much preparation would be allowed, and whether bots could be made to interact visually with a screen rather than directly through an API. His expectation that the agents would probably be superhuman was a prediction, not a tournament result.

A host said TurboPuffer had surpassed $100 million in annual recurring revenue (ARR) with roughly 50 employees. He said the company builds a vector database and search products where fast knowledge retrieval matters more than the speed of a long inference query. The host described how the team addressed network latency: traffic went through Seattle to AWS and back to servers on Google Cloud, producing a 15 ms round trip. An employee found a private internet exchange and bought a direct fiber connection, reducing the round trip to 8 ms—a 7 ms improvement. The host also recounted a Colossus profile of TurboPuffer: co-founder Simon called investor pressure to explain in advance how the company would become a unicorn “ambition bullying.” According to the host, investor Lockie Groom accepted the founders’ plan to spend a year seeking product-market fit and shut the company down if it was not working. Groom invested $700,000 at a $14 million valuation.

The hosts discussed a report that Instinct is raising money at a $30 billion valuation. One cited a progression of valuations—$100 million, $500 million, $2.5 billion, $10 billion and $30 billion—in less than a year; the discussion presented these figures and the current round as reported information, not independently verified facts. The participants stressed that few people have reliable information about the business, so outside assessments of its growth and prospects often reflect observers’ impressions, while investors make decisions based on the information available to them. A host said people they had spoken with described Instinct’s growth charts as remarkable despite competition from major AI companies and other products. He argued that launching new capabilities does not by itself guarantee adoption: promotion and placement on app screens matter. This was his reasoning about user acquisition, not a measured finding about marketing’s impact.

The hosts compared Instinct with Cursor and discussed whether a high valuation is justified amid strong competition from AI labs, which may also be suppliers. They noted that comparison is complicated because Instinct has no app, making its popularity difficult to compare directly with App Store rankings. They also considered whether large companies might acquire startups of this kind: one host argued that buying Instinct could increase competition, while acknowledging he could not predict the position of antitrust authorities. Another suggested that deals worth more than $10 billion require exceptional founder control and conviction, and doubted that leaders at Amazon or Apple would make such a move. These were the participants’ opinions and speculation, not reports of a pending acquisition.

Bloomberg’s Mark Gurman said some iPhone Duo screens crack during normal use: while the phone is being opened, a hinge component can become trapped beneath the display and puncture it. Gurman said Apple described the number of affected users as small, but the exact number of devices is unknown; he suggested it could be tens of thousands. After asking owners to send evidence, he received photos, videos, receipts and correspondence with support. Apple responded 18 minutes after he published his report.

Gurman also described difficulties getting replacement devices because stores lacked inventory. Separately, he said an acquaintance had a nonworking camera and a diagnostic message indicating it had been replaced with a non-genuine part, although the acquaintance had not replaced it. Gurman said Apple was working through the problems, but characterized the process of resolving them from the outset as a “dumpster fire.” He said his own phone had not developed cracks and hoped the issue would be resolved.

According to Mark Gurman, the announced MacBook Pro looks much the same but is lighter: the earlier model shown weighed 4.7 pounds, while the new one weighs 4 pounds. The chassis is also a couple of centimeters thinner. New features include a touchscreen, an OLED display and Dynamic Island. The new models use the same M5 Pro and M5 Max chips as the March versions.

Gurman said he bought the base M5 Pro configuration because he did not need the highest-end performance for local inference. He called the touchscreen an expected but not innovative addition and predicted Mac users would ultimately use it little because they are accustomed to a keyboard and trackpad. He considered OLED a more compelling reason to upgrade. According to the discussion, top configurations could cost about $14,600 for the MacBook Pro and more than $22,000 for the Mac Studio. Gurman also predicted M7 Pro and M7 Max versions at the end of next year.

Mark Gurman said early iPhone Duo reviews criticized app support too soon. Disney+, Instagram and YouTube, he said, supported the device from day one, and less than a week after launch only a few apps still felt awkward. Gurman attributed the initial delays to the SDK for iPhone Duo, which he said was released roughly one or two weeks before the phone went on sale.

He also described the camera as useful for FaceTime calls and said Touch ID had not been a problem for him: in less than a week, he had become accustomed to the sensor and enrolled multiple fingers twice. But Gurman pointed to software shortcomings: FaceTime picture-in-picture and quick replies to messages from notifications did not work on the inner screen. He hoped Apple would fix them in a software update. Gurman said he used the inner screen about 75% of the time and found the outer screen perfectly usable for quick interactions.