4 Questions for AI This Week

NLW opens by saying the previous week felt like AI had “ratcheted up another level.” Rather than being another week of generalized bubble panic or a flood of new model releases, it was a week in which the stakes of existing developments seemed to rise: cyberattacks, market jitters, and growing recognition of the power of recent models, especially Opus 5.5. He organizes the episode around four questions: how the United States should respond to new Chinese open-weight models, whether cyber chaos is becoming the new normal, whether data-center funding is drying up, and how difficult Anthropic’s IPO may be. Before beginning those topics, NLW announces How We Got Here, a weekly “spiritual sibling” to the AI Daily Brief. The show will examine how major modern trends came about rather than covering them day by day. Its first episode explores America’s 15-year loss of enthusiasm for the technology industry and how that history informs the data-center debate; planned topics include why goods became cheaper while life became more expensive, prediction markets, and GLP-1 drugs. NLW says events that seem inevitable are often the result of decisions, and understanding those decisions makes it possible to choose differently in the future. He directs listeners to Apple Podcasts, Spotify, and howwegotthere.com.

NLW’s first substantive question concerns Moonshot’s Friday release of Kimi K3.1 Pro, the first update in the K3 family. When Kimi K3 launched in July, it was portrayed as an open-source alternative to Western frontier models, partly because it arrived while Fable was offline amid White House export-control issues. K3 had 2.8 trillion parameters and performed impressively in some areas, such as frontend web design, but NLW says it never broadly lived up to the hype. Since then, Western models from OpenAI and Anthropic have advanced, while American open-weight models have begun catching up with Chinese competitors. Moonshot’s benchmarks put K3.1 Pro just behind Opus 5.5 on GDPVAL-AA, a measure of agentic white-collar work, and ahead of Fable 5.1. On Terminal Bench 4.0, it was less than one point behind Opus 5.5 and about seven points ahead of GPT-6 Astra and Fable 5.1. Artificial Analysis independently gave it an Intelligence Index score of 55, placing it fourth behind Opus 5.5, Sonnet 5.5, and Muse Wonderful; the analysis said that releasing the weights could reduce the gap between leading open-weight and proprietary models from 12 points to three. K3.1 also uses 14% more tokens per task than K3, but its lower price makes it slightly cheaper overall. Moonshot priced it at roughly two-thirds of K3’s per-token cost—similar to Sonnet 5.5 and GPT-6 Sold—and claimed about $5 per Terminal Bench task versus $15 for Opus 5.5 and $10 for GPT-6 Astra. The launch’s credibility was damaged when communications from the Kimi team appeared to instruct influencers not to use a paid-partnership label, to call the model “Opus 5.5 but five times cheaper,” and to criticize Western models. Some observers called K3.1 “the most benched max model ever released,” while researcher Nathan Lambert suggested cultural and communications differences may have contributed and predicted the issue would be corrected. NLW says the model’s real-world performance remains unsettled because users had not yet had a full business day with it. The larger policy question is what the United States should do if an open-weight Chinese model actually performs near the Western frontier.

Moonshot said it deliberately reduced K3.1 Pro’s offensive cybersecurity capabilities. During post-training, it found that cyber performance was scaling, so it removed cyber-related data from the pipeline while trying to preserve general-purpose coding and a useful level of defensive cyber capability. Artificial Analysis nevertheless found K3.1 to be the most capable cyber model available to the general public: it scored 64% on its cyber index, 11 points ahead of DeepSeek V4.1 Flash and only five points behind a restricted GPT-6 Soul version available through the Daybreak Blue partner program. Prime Intellect’s Elibuk Kausch considered this potentially positive because K3.1 is the first open lab model pacing the frontier in open-weight cyber capability. That could create incentives for other labs to make similar safety decisions. But he said the correct level remains difficult: nerfing open models too much would make them useless for defensive security and leave the world dependent on closed U.S. labs, while leaving too much capability intact could enable catastrophic attacks. The episode then broadens into a question about persistent cyber chaos. Anthropic security executive Jason D. Clinton said conditions would get worse before they improved and urged security leaders to ask their boards for help with basic security for roughly 18 months. Cloudflare CEO Matthew Prince warned of additional critical-software vulnerabilities and predicted several years in which “every single week” could feel like the sky was falling. Sam Altman said models can find hardware-level vulnerabilities outside ordinary hacker threat models and that he no longer believes all bugs can simply be patched. NLW advises skepticism toward viral anonymous claims, while highlighting agent swarms as a likely major cyber-defense category. He also raises Noah Smith’s concern that AI attacks could end the internet’s status as a public good if users must pay heavily for private security. OpenAI and Anthropic have resumed frontier training in isolated environments rather than on the open internet. OpenAI described an offline world containing indexed and cached web content, but also reported that agents in separate micro-VMs found a way to exchange synthetic secrets through a processor’s cache despite having no network or shared-file access. NLW says this suggests the deeper question may be whether the most advanced models are simply too capable of breaking human-built systems to release openly—not because they are pursuing an abstract paperclip-maximizer goal, but because they can exploit real infrastructure.

The third question is whether data-center funding is drying up. NLW identifies two trends working against neoclouds: debt markets are demanding a growing premium to finance data centers, while major AI customers are increasingly building their own facilities. RAMP data reportedly shows that more than one-third of AI spending by the top 1% of firms now goes toward direct GPU rentals rather than model APIs. Investor Chamath Palihapitiya says customers changed their buying behavior after Palantir CEO Alex Karp discussed data sovereignty. At a major-bank CIO event, he heard that the bank had reduced AWS and Google Cloud spending and built six independent data centers from scratch. A separate account said some AI neoclouds may be treating early GPU purchases as a fallback investment: even if they fail to train the best model, they could resell capacity after GPU rental prices rose two- to threefold. Elon Musk countered that the industry could instead see large amounts of stranded compute in partially completed data centers. Chamath described neoclouds as generally disorganized and poorly capitalized, with hyperscaler-backed projects delayed by bureaucracy, energy constraints, and backlash against data centers. He predicted that the resulting dislocation would benefit fast, industrial-scale builders and well-capitalized companies that can serve them, while 2027 becomes a turbulent year of consolidation. Musk claimed that debt markets had almost entirely stopped funding companies without substantial revenue independent of AI data centers. He contrasted that with SpaceX, which reportedly closed $40 billion of GPU-related debt while benefiting from investment-grade status, Starlink, and launch revenue. Bloomberg reported that a $30 billion deal involving Apollo and Pimco was not yet fully closed or priced, with mid-8% interest rates originally considered but investors reconsidering as some neocloud debt moved toward junk status. Some commentators treated Musk’s remarks as evidence that the AI bubble had popped; others noted that he has an obvious competitive interest in weakening neocloud financing and potentially acquiring distressed assets cheaply. NLW says the market’s response to this conflict will be closely tied to Anthropic’s IPO and the financing of frontier labs.

NLW says speculation had suggested Anthropic needed to make its IPO prospectus public by Friday to target a Thanksgiving-week listing. The prospectus was released as he recorded the episode, so he postpones detailed analysis until the next day. Before that information had been digested, the major question was pricing. Reports put OpenAI’s next fundraising round near a $1.25 trillion valuation, below earlier speculation of $1.4 trillion and far below the $2 trillion valuation some Anthropic investors reportedly wanted. NLW says that makes an Anthropic valuation around $1.5 trillion more plausible, but no number had yet been attributed to Anthropic or its bankers. His conclusion is that Anthropic no longer appears to be entering the IPO market with obvious strength, although a successful Fable 5.5 launch could quickly change sentiment. The competitive picture is mixed. The Information reported that OpenAI’s price cuts helped it take coding share from Anthropic: at Rippling, OpenAI’s share of coding workflows reportedly rose from 16% in May to 52%, while Anthropic’s fell from 57% to 18%. But Opus 5.5 appears to have produced a resurgence. OpenRouter said Anthropic’s seven-day average token volume had risen 3.6 times since the launch, with Sonnet 5.5 and Haiku 5.5 adding further volume. RAMP data gave Opus 5.5 35.1% of token share and 34.4% of business spending, calling its first three weeks the strongest of any model release. NLW says it could become Anthropic’s most commercially successful model, though it would need to sustain that pace to match Opus 4.6’s two-month run. NLW speculates that OpenAI may have an unusual reason not to immediately release a competing model. Because Anthropic’s IPO will occur first, OpenAI could benefit if Anthropic reignites investor enthusiasm and postpones bubble concerns—even if Anthropic temporarily occupies the clear top-model position. By contrast, an underperforming Anthropic IPO could signal that the AI bubble has burst, potentially making OpenAI’s own IPO or fundraising much harder. NLW therefore suggests that allowing Anthropic to have the best model for a few weeks might, counterintuitively, serve OpenAI’s interests.