Video summary

The Next Wave of Enterprise AI

Main summary

Key takeaways

News and Commentary

Summary of the Video (AI Daily Brief)

1) Trump AI executive order: voluntary cyber model sharing, but political fight over “licensing” implications

The segment opens with commentary on the “latest Trump AI executive order,” described as an unusual and confusing policy process due to internal political tension within Trump’s coalition—especially criticism from figures like DeSantis and Steve Bannon, who reportedly dislike Trump’s closeness to Big Tech.

Trigger for the renewed policy push

  • Alleged cyber capabilities in Anthropic’s advanced model work (notably “Methuselah”)
  • This allegedly became the catalyst for talk about requiring access to frontier models before release.

Timeline and changes

  • A draft circulating weeks earlier would have required safety testing and included a 90-day pre-release window for government access/testing.
  • Hours before a planned signing, Trump pulled the order, citing concerns about aspects of the policy and worrying it could slow U.S. AI progress relative to China.
  • Reporting says David Sacks intervened at the last minute to delay signing.
  • The final signed version is described as substantially similar to the earlier draft, but with a key change:
    • encouragement/incentive to share models 30 days prior to public release instead of 90.

Safety/testing vs. “licensing” debate

  • Safety testing is voluntary in both versions.
  • The host notes major labs are already inclined to submit advanced models, suggesting the “voluntary vs. compulsory” distinction may matter less in practice.
  • The new order includes language explicitly disclaiming any move toward:
    • mandatory government licensing
    • pre-clearance
    • permitting
  • Critics argue it still effectively sets up infrastructure for future regulatory control.

Government structure

  • The NSA is assigned primary responsibility for model testing, with support from other cyber/defense agencies.
  • The Treasury runs a cybersecurity “clearinghouse” in consultation with the NSA and DHS-related bodies.
  • Civilian and military agencies are instructed to harden systems against AI-enabled cyber risk.

How it was received

  • The New York Times is criticized (by OSTP) for framing it as a shift away from a “hands-off” approach.
  • David Sacks says coverage is meant for models that represent a “meaningful step change” in cyber capabilities—not incremental improvements.
  • Dean Ball and others worry about:
    • classified thresholds
    • “mission creep” toward an eventual licensing regime
  • Steve Bannon and Bernie Sanders are cited as agreeing in direction (more regulation), with Sanders arguing the EO is voluntary and weak.

2) Anthropic “Glasswing” / Mythos rollout: broader access, cybersecurity stakes, and confusing “when general access” messaging

The segment pivots to Anthropic’s context driving the EO discussion.

Expanded access via Project Glasswing

  • Anthropic expands access to the Mythos model through Project Glasswing.
  • Adds 150 new partners across 15 countries.

Broader sector coverage + cybersecurity stakes

  • The rollout expands into sectors such as:
    • energy, water, communications, health care, and computer hardware
  • The argument: a successful cyber attack on partner codebases could have catastrophic impacts, potentially affecting 100+ million people.

General access plans are uncertain

  • Anthropic previously suggested Mythos-level capability could reach general access within weeks.
  • In the update, Anthropic says they are moving quickly but need “highly robust safeguards” that they claim they (and other AI developers) don’t yet have.
  • The host interprets this as a walking back of earlier timelines.

Practical constraints: expensive to test

  • The Information reports Mythos testing is extremely costly.
  • Teams burn millions of dollars’ worth of tokens quickly.
  • Anthropic subsidizes early usage, but firms still plan budgets around Mythos for strategic reasons.

3) AI supply chain: SK Hynix plans major memory capacity expansion due to token-driven demand

On hardware constraints, the host reports:

  • SK Hynix plans to double memory chip manufacturing capacity by the end of the decade to relieve AI server memory shortages.
  • The token boom is causing shortages.
  • High-bandwidth memory costs have reportedly more than doubled this year.
  • Even with investment, the shortage could last until 2030, limiting near-term relief.
  • Long-term investment is framed as necessary for sustainability.

Main Episode: “The Next Wave of Enterprise AI” = Knowledge-Work Interfaces + Agentic Workloads + Cost Management

4) OpenAI’s enterprise/knowledge-work push with Codex: “strange abundance” and workflow changes

The show frames enterprise AI’s next phase as moving from a “subsidy era” to a “scarcity/token shortage era,” driven by the growth of agentic workloads that consume far more compute/tokens.

Adoption isn’t just cost—it’s “use-case fluency”

  • Beyond price, adoption depends on:
    • use-case fluency
    • new interfaces

Codex usage trend

  • Google searches for “Codex” reportedly surpassed “Claude, Code” for the first time (per The Information).
  • Codex reportedly reaches 5 million weekly active users.
  • Biggest growth comes from non-technical knowledge workers, adopting Codex 3x faster than developers.

Design thesis from OpenAI’s report (“Next Era of Knowledge Work”)

Knowledge work is slowed by three frictions:

  1. finding relevant inputs across messy systems
  2. coordinating information
  3. approvals and verification

While workers can produce artifacts faster than ever, they still waste time managing, reconciling, and moving information.

Codex is positioned as a “factory redesign” for knowledge work—shifting from sequential drafting to orchestrating multiple tasks.

Observed behavior shift: more parallel work

  • Users increasingly run parallel Codex tasks.
  • About 50% do more than one at a time (up from under a third in mid-April).
  • This is described as enabling one person to scale like a small team.

Codex feature highlights

  • Annotations: interact precisely with document context (selecting exact regions instead of describing in words).

  • Role-specific plugins: bundles of apps/skills/workflows for domains like:

    • sales, data analytics, creative production, product design, investing, investment banking These are framed as lowering setup burden and “productizing” best practices.

  • Sites: turn Codex-created artifacts into shareable websites/web apps.

    • Positioned as “cloud artifacts but on steroids”
    • Makes interactive sharing easier and more secure
    • The host compares this to a knowledge-work primitive (like documents/spreadsheets), but with disposable, purpose-built web experiences.

5) Cost management becomes a central enterprise strategy: token caps and cheaper model usage

Organizational cost control example

  • Uber reportedly imposes a $1,500 monthly token spending cap per employee.

Microsoft’s parallel theme: model variety and cost positioning

At Build, Microsoft releases seven AI models:

  • Image 2.5 / Image 2.5 Flash
  • Transcribe 1.5
  • Thinking 1
  • Voice 2 / Voice 2 Flash
  • Code 1 Flash

The flagship is MAI Thinking 1:

  • A claimed 1T-parameter mixture-of-experts model aimed at reasoning/agentic/tool use.

Mixed public reaction on benchmarks

  • Some commentators question competitiveness based on test results.
  • Others argue Microsoft is playing a different game:
    • not just raw performance
    • but cost-efficient frontier tuning and deployment

Microsoft’s “frontier tuning” strategy

  • Mustafa Suleyman: early adopters see improved win rates at 10x lower cost on specific tasks (described as McKinsey-like) via tuned models.
  • Satya Nadella emphasizes the shift from “consuming” frontier models to participating in the frontier ecosystem.

Host’s concluding outlook

  • The next phase (especially the second half of 2026) will focus on making newly unlocked enterprise AI capabilities work affordably, not just making them available.

Presenters / Contributors

  • Host / narrator: AI Daily Brief (unnamed)
  • Mentioned contributors/people:
    • David Sacks
    • Dean Ball
    • Steve Bannon
    • Bernie Sanders
    • David Remler (Center for a New American Security)
    • Simon Smith (ClickHealth)
    • Shawn Wang
    • Ali Bakhouch (Prime Intellect)
    • Ethan Mollick
    • Mustafa Suleyman
    • Satya Nadella
    • Chey Tae-won / Shea (SK Hynix leadership as cited in the segment)
    • The New York Times
    • The Information
    • McKinsey

Original video