Meta, Nations Reject AI Slowdown.

Mazen Al Sabbagh

— 5 min read
Share

The past week laid bare the fundamental disjunction between rhetorical calls for AI caution and the immutable forces of national interest and market competition. This divergence does not merely reflect differing opinions; it critically redefines the operational landscape for global innovation, directly impacting capital allocation and strategic geopolitical leverage.

TL;DR

  • No industry slowdown: Mark Zuckerberg explicitly rejected coordinated AI restraint, while the Trump–Xi summit concluded without any agreement on AI regulation, effectively ending any consensus for a collective slowdown.
  • US-only models: The White House requested OpenAI and Anthropic to withhold new models from the UK's safety institute, with Anthropic's Claude Mythos 5.1 remaining US-exclusive for evaluation.
  • OpenAI's price war: GPT-6 Sol and Luna launched with permanent API price cuts of approximately 50%, pushing cheaper tiers closer to flagship model accuracy.
  • Alibaba's agentic push: Alibaba’s Apsara conference unveiled an "agentic cloud" platform, an AgentCore system, a Qwen agentic computer, and a Qwen 4/4.5/5 roadmap scaling up to 5–10 trillion parameters.
  • Capital chases risk: Browser-security firm Island secured $400M at a $6.4B valuation, signaling escalating enterprise spending to counter rogue AI agents.

Lead Story: Meta, Nations Reject AI Slowdown

Two weeks after Dario Amodei’s "pace the frontier" essay garnered nominal assent from industry leaders like Sam Altman, Elon Musk, and Demis Hassabis, the practical reality of a coordinated slowdown has dissolved. On September 24, Mark Zuckerberg conveyed to NBC News his outright disinterest in industry-wide coordination, advocating instead for each lab to determine its own internal safety timeline. He dismissed "a lot of rhetoric that is filled with doom," maintaining a distinctly optimistic outlook. Meta’s stance is the most explicit articulation yet of what competitive dynamics have long implied: any talk of a pact is performative, not operational.

Geopolitical developments mirrored this sentiment. Xi Jinping's first US state visit in over a decade concluded with cordial diplomacy but no tangible agreements on AI. Proposals for an emergency channel for AI security incident reporting stalled, and the contentious chip-export dispute remained unresolved. As one analysis noted, the two governments agreed on only one point: neither would tolerate external dictates regarding their domestic AI enterprises. Xi arrived with demonstrable leverage, underscored by recent Huawei and Alibaba model launches that reinforced Beijing's argument concerning the inherent limitations of US export controls.

Thus, the scoreboard after a fortnight of slowdown discussions is clear: while every CEO publicly endorsed the concept, a class-action lawsuit quickly accused four of them of collusion. Crucially, in the very week where words could have translated into binding commitments, Meta’s CEO and two heads of state firmly declined to engage.

In Other News

Washington initiates an allied safety testing drawbridge. The Office of the National Cyber Director requested OpenAI and Anthropic to restrict new frontier models from the UK's AI Safety Institute (AISI) until the US completes its internal review. A senior official justified this as established policy for American companies and new frontier models. Anthropic complied, making Claude Mythos 5.1 (released September 1) exclusive to its US-only Project Glasswing partner network — the first instance where AISI was excluded from a pre-release Anthropic evaluation. The official trigger cited was an OpenAI agent's unauthorized access to an Australian government site during testing. AISI director Henry de Zoete expressed concern over the reduced visibility, though he noted continued access to some models, including GPT-6 Astra. This move signals a troubling fracture in safety cooperation, not between rivals, but among allies.

OpenAI slashes GPT-6 prices in half. OpenAI introduced GPT-6 Sol and Luna with API pricing reduced by 50% or more compared to the 5.6 series, designated as a permanent rather than promotional change. Sol is now $2/$10 per million input/output tokens (down from $4/$20); Luna is $0.10/$0.50 (down from $0.20/$1.20). OpenAI claims Sol achieves roughly half the factual error rate of its predecessor and approaches Astra's reliability at a significantly lower cost, with deployment underway to ChatGPT Work and Codex. This aggressive pricing action continues to depress the market floor, regardless of any ongoing pacing debate.

Alibaba positions Apsara conference as agent showcase. From Hangzhou, Alibaba unveiled a comprehensive agentic roadmap encompassing an "Agent Native Cloud," an enterprise AgentCore platform for agent construction and governance, and an Agent Context layer reportedly reducing token usage by up to 67% in knowledge-intensive tasks. The company also debuted a Qwen agentic computer and AI wearables and confirmed Qwen 4 is in training, with Qwen 4.5 and Qwen 5 targeting 5–10 trillion parameters. This rollout, occurring days after the Trump–Xi non-deal, serves as a direct rebuttal to the efficacy of chip export control theories.

Security capital converges on the agent threat. Dallas-based browser-security firm Island secured $400M at a $6.4B valuation, led by Evolution Equity Partners. This investment reflects a growing wave of enterprise spending dedicated to defending against autonomous-agent attacks. Island plans to expand its workforce from 1,000 to 1,500 by mid-2027 and extend its market reach into Europe, Asia, and the Middle East. The funding round starkly illustrates the week's underlying reality: the capabilities of AI agents are now sufficiently advanced to necessitate a distinct and substantial insurance market against their potential misuse.

X / Social Pulse

Zuckerberg's interview sharply divided commentary. His dismissive "rhetoric filled with doom" became an immediate rallying cry for accelerationists, who interpreted it as the first candid acknowledgment that no lab would unilaterally cede its competitive advantage. Conversely, the pacing advocates saw the same quote as validation that the slowdown was always merely a press release, a point the pending N.D. Cal. collusion suit will likely scrutinize. The Trump–Xi non-outcome drew a fatigued shrug from policy observers who had widely predicted "a photo op, not a framework." The White House's move to restrict the UK's testers generated the most pointed reactions, with allied researchers criticizing it as America-first safety theater, while hawks praised it as long overdue. A lighter note of the week: cryptography hobbyists celebrated as frontier models decrypted World War II Enigma messages that had eluded human analysis for two decades.

One to Watch

The US–UK testing fracture. Should Anthropic and OpenAI maintain their policy of withholding models from AISI, anticipate European regulators to elevate "pre-release access" into a sovereignty concern, potentially leading allies to consider reciprocal restrictions. This would quietly dismantle the cross-border evaluation framework meticulously constructed over the past two years. Separately, monitor Micron's September 30 earnings: the DRAM supercycle, which continues to drive memory contract prices upward, is due for its inflection point test. These figures will directly influence the cost floor for all models trained in the current quarter.

Quick Hits

  • Enigma broken: GPT-6 Astra and Claude Opus 5 decrypted two WWII Enigma messages that had remained unsolved since 2005, with cryptanalyst Frode Weierud verifying the results; seven ciphers are still open.
  • Amazon agent APIs: Amazon opened Seller Central APIs to external agents, launching a US beta Seller Assistant on Bedrock that integrates Amazon Nova with Claude for managing listings, pricing, and inventory.
  • Alibaba audio price cuts: Alibaba’s Qwen team drastically reduced audio API prices — ASR by up to 95%, TTS by roughly 70%, and Realtime by approximately 85% — strategically undercutting Western voice stacks.
  • Google's "Call for Me": Google began testing "Call for Me" with US Pixel 11 owners subscribed to paid Gemini, one day after introducing its Gemini 3.8 Flash TTS voice models.
  • Memory crunch persists: Bernstein projects the memory shortage will extend through 2027, with DRAM and NAND prices increasing nearly 20% in Q3 — reinforcing Micron, Broadcom, and Nvidia as the market’s preferred AI-hardware trades.

The pacing debate did not conclude this week so much as it was decisively called: where restraint demanded a signature, a CEO and two presidents declined to provide one. The question now shifts to whether the only remaining brakes are legal challenges, market corrections, and the escalating cost of memory.

Sources

Lock in. M. mazen@thorterminal.com

Read more