Aaron Levie on Why Open AI Wins
Theo Jaffee, Sofia Puccini, Aaron Levie
AI-assisted summary of Aaron Levie on Why Open AI Wins — original source. Check the original for full context.
Source published:
Key ideas
- Box CEO Aaron Levie argues that open-weight AI models expand total enterprise adoption without cannibalizing frontier labs, while calling for high-net-worth philanthropists to endow a $10 billion domestic non-profit open-weight lab. (0:00)
- Levie warns that heavy-handed U.S. export controls and protectionism will fail to slow Chinese AI progress, instead incentivizing the creation of an independent foreign hardware and software stack that permanently displaces Western platforms. (6:38)
- Enterprise software economics are shifting toward dynamic model routing, as companies avoid single-vendor lock-in and capture outsized value at the workflow, security, and context layer. (26:15)
- Rather than eliminating engineering jobs, AI tooling at Box has dramatically expanded software roadmaps by reducing multi-year initiatives and unlocking previously neglected backlog items. (22:30)
Detailed notes
## Background Published on September 5, 2026, on the MTS podcast by hosts Theo Jaffee and Sofia Puccini, this episode features Box co-founder and CEO Aaron Levie making the strategic case for open-weight artificial intelligence. According to the show notes, the discussion explores the economics of open versus closed models, geopolitical AI competition with China, the evolution of software engineering, and why the enterprise routing layer will capture lasting value.
## TL;DR Box CEO Aaron Levie argues that open-weight AI models do not cannibalize closed frontier labs, but instead drive broad market diffusion, accelerate proprietary innovation cycles, and settle margins sustainably around compute infrastructure. Levie pushes back against national security hawkishness that seeks to isolate Chinese models, warning that export bans will only catalyze independent foreign silicon ecosystems while burdening American developers with artificially expensive proprietary APIs. He also refutes the idea that AI will eliminate software engineers, explaining that productivity gains actually expand product ambition and unlock previously shelved backlogs. Finally, Levie predicts that vendor lock-in will fail in the enterprise, ensuring that durable value accrues to the applied routing layer that bridges enterprise data, permissions, and evolving models.
## Atmosphere & style The conversation moves at an energetic, high-bandwidth clip, blending insider Silicon Valley humor with rigorous strategic and economic analysis. Levie, Jaffee, and Puccini share an easy, self-aware rapport, effortlessly weaving between tongue-in-cheek commentary on astrology image generators and serious critiques of regulatory guardrails and national security policy. It feels like an authentic, behind-the-scenes strategy whiteboard session—candid, pragmatically grounded in enterprise customer realities, and unapologetically dismissive of corporate protectionism.
## Bulletin - `00:00` Box CEO Aaron Levie argues that open-weight AI models expand total enterprise adoption without cannibalizing frontier labs, while calling for high-net-worth philanthropists to endow a $10 billion domestic non-profit open-weight lab. - `06:38` Levie warns that heavy-handed U.S. export controls and protectionism will fail to slow Chinese AI progress, instead incentivizing the creation of an independent foreign hardware and software stack that permanently displaces Western platforms. - `26:15` Enterprise software economics are shifting toward dynamic model routing, as companies avoid single-vendor lock-in and capture outsized value at the workflow, security, and context layer. - `22:30` Rather than eliminating engineering jobs, AI tooling at Box has dramatically expanded software roadmaps by reducing multi-year initiatives and unlocking previously neglected backlog items.
## Key takeaways - `00:00` Open-weight models foster broader AI adoption by enabling deep customization and specialized workloads, which in turn accelerates proprietary frontier lab innovation. - `03:36` Model distillation does not cross ethical or legal boundaries, especially given that frontier labs train on public web data and directly collect API revenue from distillation workloads. - `03:36` Synthetic code and reasoning outputs proliferate across public repositories, making the strict policing of downstream model distillation practically impossible. - `06:38` Restricting access to international open models risks saddling the U.S. with artificially high compute costs while competitors leverage free, hyper-efficient open architectures. - `06:38` Export controls will not halt Chinese AI advancements due to China's deep computer science talent pool, state-backed industrial base, and massive capacity for data curation. - `10:00` The primary economic driver of enterprise AI is compute rather than model secrecy, as inference costs dictate where revenue ultimately settles. - `10:00` Frontier labs will see model gross margins naturally compress to a sustainable 20% to 40% over bare compute due to intense multi-lab competition. - `10:00` U.S. frontier labs should consider open-sourcing prior-generation (n-1) models to preserve developer mindshare while monetizing the resulting cloud infrastructure workloads. - `15:00` Anthropic avoids open weights primarily due to short-term API monetization advantages and a founding ideology centered on centralized oversight and guardrail enforcement. - `15:00` Claude 3.5 Opus delivers strong price-to-performance gains for unstructured document analysis across regulated enterprise sectors like finance, law, and life sciences. - `20:00` Premature, aggressive safety guardrails and silent model downgrades risk rendering frontier models unusable for complex enterprise tasks like permissions auditing. - `22:30` AI expands the scope of engineering ambition: companies claiming AI eliminates the need for software engineers likely lack ambitious product roadmaps. - `25:00` Autonomous coding agents frequently overestimate task timelines because their pre-AI training data reflects cautious human engineer communication habits. - `26:15` The applied AI routing layer captures durable value by connecting enterprise permissions, context, and operational workflows across interchangeable foundation models.
## Detailed walkthrough ### `00:00` Introduction and the Jensen Huang Open-Weights Letter
MTS hosts Theo Jaffee and Sofia Puccini introduce Box co-founder and CEO Aaron Levie, highlighting Box's footprint across cloud content management, enterprise documents, security permissions, collaboration, and enterprise AI integrations, as well as Levie's prolific presence on X (formerly Twitter) under the handle `@levie`. Before diving into heavy policy, the hosts open with a tongue-in-cheek detour into breaking industry gossip regarding a supposed massive acquisition in the "astrology image gen space." When Puccini asks if astrology will embrace open-source architecture, Levie deadpans that astrology might remain closed source for the time being, prompting Jaffee to joke that the ecosystem will need to organize against proprietary horoscopes.
Turning to serious matters, Puccini brings up an open letter advocating for open-weight AI models, organized by NVIDIA CEO Jensen Huang and signed by enterprise tech leaders, including Box. Levie treats the letter as an industry Rorschach test, noting that signatories and observers project their own worldviews onto its contents, but explains that Box signed it based on two core arguments. First, open-weight models do not cannibalize closed frontier labs in a zero-sum contest; instead, they dramatically broaden the aperture of AI adoption by letting developers, startups, and enterprises customize, fine-tune, and build specialized workloads on top of foundational architectures. That expansion of total use cases ultimately pushes proprietary frontier labs to accelerate their own innovation cycles. Second, Levie interprets the statement as an urgent domestic call to arms: rather than merely panicking over rapid advances in Chinese open models or treating model distillation as an illicit shortcut, the United States must actively fund and build its own world-class open-weight ecosystem.
Levie readily concedes that the safety lobby raises genuine arguments regarding catastrophic capability overhangs in future iterations of frontier models. However, he maintains that stifling open weights out of precautionary fears will simply starve the broader software economy of vital infrastructure. Defending the letter's pro-diffusion thesis, Aaron Levie argues: “I almost unequivocally take the stance that more open innovation and open weights innovation is good for AI broadly and the diffusion of AI.”
```mermaid flowchart TD accTitle: Open Weights Flywheel accDescr: Diagram illustrating how open-weight AI drives adoption and pushes closed frontier labs to innovate faster.
subgraph OpenEcosystem["Open-Weights Ecosystem"] A["Open-Weight Foundations"] --> B["Developer Customization & Fine-Tuning"] B --> C["Expanded Commercial & Enterprise Use Cases"] end
subgraph MarketPressure["Market Feedback Loop"] C --> D["Broad AI Market Diffusion"] D --> E["Competitive Pressure on Frontier Labs"] E --> F["Proprietary Labs Innovate Faster"] end ```
---
`03:36` The Distillation Dilemma and the Case for an American Open Lab
The hosts push Levie on model distillation, asking where he draws the line between legitimate synthetic data optimization and improper intellectual property infringement. Levie rejects the premise that distillation crosses an inherent ethical or legal line. He points out the logical inconsistency of proprietary labs scraping the public internet without individual creators opting in, while simultaneously asserting that no subsequent model can train on their model's public or commercial outputs. While Levie sympathizes with closed-lab executives who try to enforce restrictive terms of service to defend their competitive moats, he stresses that attempting to treat output-based training as a civil or criminal violation undermines the foundational logic of machine learning.
Levie and Jaffee also highlight an ironic economic reality: frontier providers like Anthropic and OpenAI are actively monetizing distillation. Developers pay substantial API usage fees while extracting synthetic datasets, generating more direct revenue for the frontier lab than the original human creators ever received during pre-training. Furthermore, Levie offers a thought experiment regarding synthetic data diffusion: if developers query models like GPT-4 or Claude via API to write thousands of software projects, and those codebases are then pushed to public GitHub repositories, downstream models inevitably ingest that distilled data during standard web scrapes.
Looking at the broader competitive landscape, Levie dismisses the narrative that foreign rivals advance solely through illicit distillation. He emphasizes that labs in China are achieving genuine algorithmic breakthroughs independent of Western outputs. Rather than obsessing over intellectual property fencing, Levie urges American philanthropists and tech titans to fund native open-source research at scale. Levie outlines his ideal vision for domestic innovation: “The best case scenario personally would be you get a few people that are at a stage in life where they can just be like, yep, I'm going to drop $10 billion on building an open lab in America.” Jaffee and Puccini joke about tagging the Bill & Melinda Gates Foundation to underwrite the initiative, while Levie insists that an American equivalent of Moonshot AI or DeepSeek is overdue.
- Proprietary AI labs collect substantial API revenue from the very distillation workloads they publicly decry. - Organic data pollution makes policing distillation nearly impossible once synthetic text, reasoning traces, and code proliferate across public internet repositories. - Levie calls for high-net-worth American philanthropists or consortiums to endow a $10 billion non-profit open-weight research lab to counter foreign dominance.
---
`06:38` The China Threat, the Dwarkesh Debate, and Hardware Lockout
The discussion moves to the geopolitical fallout of China dominating the open-weight landscape. Jaffee outlines the standard national security hawkish critique: if American developers become addicted to Chinese open models like DeepSeek or Moonshot, Beijing could eventually pull access, dictate software norms, or undercut American enterprise software margins. Levie counters that trying to cut off international open models is as economically self-defeating as arguing an iPhone should cost $5,000 to manufacture domestically rather than leveraging mature Asian supply chains. Forcing the United States onto artificially expensive proprietary APIs while the rest of the world builds on free, hyper-efficient open weights would guarantee an American competitive disadvantage over the next decade.
Levie references the polarizing debate between Jensen Huang and Dwarkesh Patel on the *Dwarkesh Podcast*, which served as another industry flashpoint. While some tech commentators believed Patel was right to cross-examine Huang on national security risks, Levie sides firmly with Huang's analysis of technological inevitability. Levie argues that geopolitical export bans will never convince China to abandon AI, given the technology's strategic primacy. China possesses an immense base of domestic engineering talent, massive state-backed industrial fabrication capacity, and the organizational muscle to mobilize hundreds of thousands of people for synthetic data curation and reinforcement learning annotation.
Levie warns that heavy-handed U.S. export controls and protectionist postures will trigger a dangerous unintended consequence: rather than slowing Chinese AI, restrictions simply catalyze the creation of an entirely independent hardware and software stack. If foreign labs are locked out of NVIDIA's ecosystem, they will design and scale competing domestic accelerators, deploy them across sovereign clouds throughout the developing world, and permanently displace Western technology platforms. Reaffirming his agreement with the NVIDIA CEO's geopolitical calculus, Aaron Levie states: “I was watching it and just was like, yeah, Jensen's obviously right. You might not like the conclusion, you might disagree with it, but he's obviously right that the following will happen. If you block off China, China is not going to give up on AI.”
- Restricting domestic access to foreign open models risks leaving the U.S. with artificially high compute costs while global competitors build on dirt-cheap, highly capable open weights. - China's combination of elite computer science talent, deep industrial manufacturing, and massive labor pools for manual data generation ensures it will remain a frontier player regardless of chip bans. - Western hardware embargoes risk accelerating the global adoption of alternative silicon stacks across non-aligned sovereign cloud infrastructures.
`10:00` Geopolitics, Jensen Huang, and the True Economics of Open Weights
Aaron Levie pushes back against the belief that Western labs can simply stonewall China to maintain an insurmountable lead. In his view, assuming that strict isolation will permanently preserve American primacy rests on shaky logic: “So I think it's, I just don't think there's a scenario here where you can kind of close off completely to China and somehow you dramatically slow them down.” With rapid advancements such as Kimi k1.5 and similar releases, the technical gap between the US and China is visibly narrowing rather than expanding. The hosts briefly detour into Dwarkesh Patel’s recent interview with Jensen Huang, calling it “generational,” “absolute cinema,” and “tech reality TV,” specifically laughing over Jensen’s visceral rejection of defeatism: “I'm not a loser. You're not talking to someone who woke up a loser.”
```mermaid flowchart TD accTitle: Value Capture in Open versus Closed AI Architectures accDescr: Diagram illustrating how competition drives margins down to infrastructure and how labs monetize open weights.
subgraph Competition ["Market Realities"] A["3–5 Leading US Frontier Labs"] --> B["Token Prices Converge Toward Compute Costs"] B --> C["Inference Margins Settle at 20% to 40%"] end
subgraph Monetization ["Monetizing Open-Weight Ecosystems"] C --> D["Provide Preferred Cloud Infrastructure"] C --> E["Post-Training & Fine-Tuning Environments"] C --> F["Trusted Enterprise Brand & Safe Deployment"] end
subgraph UltimateSpend ["Capital Flow"] D & E & F --> G["Revenue Accrues to High-End GPU Clusters"] end ```
When asked what trade-offs the US faces by openly releasing capable frontier-level weights, Levie explains that the primary driver of enterprise software value is compute, not model secrecy. Even if an architecture is open-sourced, the money ultimately flows directly to the infrastructure stack because inference costs dominate. Levie rejects the zero-margin thesis, estimating that competition among three to five major US labs will naturally compress model gross margins from sky-high software levels (70%–90%) down to a sustainable 20% to 40% over bare compute. Under those economics, hosting open models captures nearly the same bottom-line value as running closed proprietary APIs.
- **Market lead assumptions:** Theoretical arguments for shutting down open weights depend on the flawed premise that Western labs enjoy an unbridgeable, accelerating technical moat over Chinese competitors. - **Inference as the profit center:** Real money in artificial intelligence accrues to inference compute; running open models still demands vast GPU infrastructure that laptops and local edge devices cannot replace. - **Fast-following with open weights:** Levie argues frontier labs like OpenAI should systematically open-source their prior-generation models (n-1), keeping developer mindshare within their model families while monetizing the resulting infrastructure workloads.
`15:00` Why Frontier Labs Stay Closed and Claude Opus 3.5 Benchmarks at Box
Addressing why dominant providers like Anthropic and Google avoid open-sourcing previous flagship architectures, Levie identifies two key friction points: temporal incentives and safety ideology. Financially, maintaining an exclusive closed API generates superior cash flows in the short run: “Because tomorrow, if I release a closed model, I just will literally make more money if I just keep it closed and I force everything to route through my API.” Over a multi-year horizon, however, commodity pressures level out those structural advantages. Culturally, organizations like Anthropic view open weights as an unacceptable safety hazard; they believe a single centralized entity must oversee token streams to intercept prompt injections, monitor queries, and guarantee guardrails.
```mermaid flowchart LR accTitle: Enterprise Knowledge Work Dual Evaluation Axes accDescr: Visual mapping showing the two core capability axes required for enterprise AI models.
subgraph Axis1 ["Horizontal Intelligence"] H["General Reasoning & Large Context Synthesis"] end
subgraph Axis2 ["Vertical Expertise"] V1["Life Sciences & Bio Research"] V2["Legal Precedent & Contract Logic"] V3["Institutional Banking & Finance"] end
H & V1 & V2 & V3 --> Eval["Enterprise Benchmark Performance"] ```
The discussion transitions directly to real-world performance evaluations with the release of Claude 3.5 Opus. Drawing on Box’s production workloads handling unstructured documents for law firms, life sciences companies, and major financial institutions, Levie notes that enterprise adoption requires simultaneous gains across two distinct vectors: broad general reasoning and intense domain-specific comprehension. Box’s internal benchmarks demonstrate that Claude 3.5 Opus delivers massive generational leaps over prior models, particularly in synthesis and industry accuracy.
- **Anthropic's safety stance:** Anthropic’s refusal to open-weight models is rooted in core founding philosophy—open weights eliminate the central oversight needed to filter dangerous queries and manage token-level security. - **Enterprise evaluation vectors:** Box tests frontier releases against complex corporate workloads that combine high-level general intelligence with rigorous vertical knowledge in regulatory, medical, and legal contexts. - **Claude 3.5 Opus vs. competing flagships:** While models like OpenAI's GPT-4o (or internal variants referred to as "Fable") may edge out slight wins on raw qualitative vibes, Opus represents a Pareto-superior choice for enterprise knowledge work due to its massive price-to-performance advantage at roughly half the running cost.
`20:00` Anthropic’s Guardrails, Downgrades, and the Perils of Premature AI Regulation
Discussing recent frontier model evaluations, Aaron Levie reflects on internal tests comparing cutting-edge models. He notes that in their benchmark testing, Fable demonstrated superior capability over version 4.8 on difficult technical problems, delivering a level of performance that could not be offset simply by token cost advantages. When asked whether the comparison was against 4.8 or an upcoming Opus 5, Aaron explains that prior to having access to version 5, he lacked enough internal coding evaluations to measure that specific gap, particularly because many newly released benchmarks focus on general knowledge work and agentic tool use rather than pure coding.
The conversation turns to a frustrating friction point users have encountered with Fable: the tendency for the system to quietly downgrade queries to a lower-tier model like 4.8 when safety or security filters trigger. Aaron observes that developers frequently encounter this when analyzing permissions or access control code, which alarms the safety guardrails and strips away the model's raw problem-solving capability. While Aaron commends Anthropic for attempting to structure a constructive dialogue with policymakers through a multi-point risk framework, he voices deep skepticism about who determines those risk thresholds and whether the identified risks reflect reality. Aaron warns that aggressive, early throttling could choke adoption: “People will simply not use AI if this is the kind of ongoing environment.” Because of the nuance involved, he argues it is far too early to lock models down with restrictive regulations.
- Internal Box testing showed Fable significantly outperforming 4.8 on complex technical problems beyond what token economics could balance. - Automated safety triggers frequently downgrade high-capability prompts to 4.8 during access control or permission audits, depriving engineers of top-tier model intelligence. - Aaron supports risk assessment frameworks in principle but questions who defines the criteria, maintaining that premature guardrails threaten to make frontier AI unusable.
`22:30` Expanding Software Engineering Ambition and Unlocking the Backlog
When asked how advanced AI models are altering the practice of software engineering and hiring expectations at Box, Aaron rejects the notion that software engineers are becoming obsolete. Instead, he explains that AI fundamentally alters the economics and psychology of product planning. Before AI, development roadmaps were restricted to a narrow middle band: teams routinely rejected tiny tasks because spending a week on a minor issue was not worth the developer hours, while simultaneously shelving massive, multi-year initiatives because they were deemed too risky or difficult to execute.
AI collapses both ends of that spectrum. Multi-year undertakings shrink into manageable short-term efforts, while low-priority backlog bugs that previously took a week can now be knocked out in two hours. Aaron shares that Box currently has multiple dozens of active development projects that never would have received the green light in a pre-AI world. Rather than cutting technical staff, Box is actively trying to hire more software talent, constrained primarily by overall business budgeting across departments rather than a shortage of work. Challenging tech leaders who use AI as a rationale for developer layoffs, Aaron asserts: “I generally think if you think that you've kind of eliminated the need for software engineers, there's just no chance you're being ambitious enough with your product roadmap.”
```mermaid flowchart TD accTitle: Impact of AI on Software Roadmap Economics accDescr: Diagram contrasting pre-AI roadmap constraints with post-AI software scope expansion.
subgraph PRE["Pre-AI Software Planning Constraints"] direction TB A1["Tiny Fixes (~1 week)\n'Not worth developer time' -> Ignored"] A2["Middle Band Projects\nStandard, moderate feature roadmap"] A3["Multi-Year Initiatives\n'Too complex and costly' -> Rejected"] end
subgraph POST["Post-AI Roadmap Expansion"] direction TB B1["Backlog Clearance\nTiny fixes solved in ~2 hours"] B2["Expanded Core\nFaster delivery of customer value"] B3["Moonshots Unlocked\nMulti-year complexity drastically compressed"] end
A1 -.->|"Accelerated by AI"| B1 A2 -->|"Augmented capacity"| B2 A3 -.->|"Compressed by AI"| B3 ```
- Pre-AI roadmap planning stranded teams in a narrow middle band, ignoring both week-long backlog tickets and ambitious multi-year projects. - Box has dozens of active projects underway that would have been rejected as technically unfeasible or economically unjustifiable without AI. - Engineer productivity gains translate into resolving more customer problems faster, heightening demand for engineering talent rather than eliminating it.
`25:00` The Humor of AI Overestimation and Slack-Trained Timelines
The hosts and Aaron share a laugh over the quirky pacing and conservative estimates built into autonomous coding agents. Theo and Sofia note that when prompted with a task, agents often wildly overestimate the required effort, warning the user that a feature will take an entire weekend or even a month, only for the human developer to wrap it up in three hours. Aaron quips that this behavior is an inevitable consequence of the training data: “That's because it was trained on all of our Slack messages pre-AI.” He jokes that the models merely learned to mimic the classic software engineer defense mechanism of declaring every request a two-year project.
The hosts observe a similar disconnect in mathematical reasoning. When asked when an AI system might successfully disprove an open conjecture in mathematics, models typically generate conservative five-year projections—completely oblivious to the fact that AI models have already begun solving and disproving open mathematical problems in the real world.
- Coding agents regularly overestimate implementation timeframes by orders of magnitude, predicting weeks of work for three-hour jobs. - Aaron jokes that cautious timelines stem from pre-AI internal communications where human engineers reflexively inflated project estimates. - AI models display similar lagging conservatism in forecasting mathematical breakthroughs that have already occurred.
`26:15` The Rise of Model Routing and Value in the Applied AI Layer
Turning to enterprise architecture, the hosts discuss whether enterprises can remain loyal to any single frontier lab when model releases and benchmark leapfrogging occur at a breakneck pace. Aaron agrees that vendor lock-in is unviable and argues that the market is inevitably moving toward intelligent model routing. While acknowledging his inherent bias as the leader of Box—a platform operating directly in the applied AI tier—Aaron highlights that companies like Box, Cognition, Factory, Cursor, and Replit thrive in an ecosystem where diverse models excel across different axes. In this paradigm, super-frontier models handle high-order orchestration while cheap, cost-tuned workhorses process repetitive sub-tasks.
```mermaid flowchart TD accTitle: Enterprise Model Routing Architecture accDescr: How the applied enterprise layer routes between diverse foundation models based on task, cost, and workflow context.
subgraph APP["Applied AI & Orchestration Layer (Box, Cursor, Replit)"] direction TB Context["Enterprise Context: Data, Permissions & Workflows"] Router{"Intelligent Model Router"} Context --> Router end
subgraph MODELS["Underlying Foundation Models"] direction LR M1["Frontier Orchestrators\n(e.g., Fable, Opus 5)"] M2["Workhorse / Cost-Tuned\n(e.g., 4.8, GPT-5.6)"] M3["Specialized / Open Models\n(e.g., Grok 4.5, Llama)"] end
Router -->|"Complex Reasoning"| M1 Router -->|"High Volume / Low Cost"| M2 Router -->|"Specific Domain Tasks"| M3 ```
Aaron points out that with major U.S. competitors such as Google, OpenAI, xAI, and others engaged in a continuous warpath to reduce inference costs while pushing the frontier, enterprises face acute decision paralysis. The applied routing layer resolves this dilemma by abstracting the underlying models: organizations can configure their data pipelines, security controls, and operational workflows once, seamlessly toggling between Fable, GPT-5.6, or Grok 4.5 as capabilities evolve. Aaron argues that horizontal foundation labs will struggle to capture deep vertical workflows in regulated industries like legal, healthcare, and finance because raw models lack the necessary surrounding apparatus: “It needs to get access to the right data. It needs to be implemented into the workflow in the right way. And that's why I think you're going to have just a tremendous amount of value creation at that layer of the stack.”
- Enterprise AI architecture is shifting to dynamic routing across multiple providers to optimize for both raw intelligence and cost efficiency. - Applied platforms insulate customers from model churn, allowing organizations to adopt next-generation models without refactoring workflows. - Domain-specific value accrues to the application layer because foundation models require deep contextual data, access controls, and workflow integration that horizontal labs cannot provide out of the box.
`29:30` Rapid News Cycles and Outro
Wrapping up the interview, the hosts and Aaron reflect on the dizzying speed of the AI news cycle over the preceding week. The hosts marvel at how quickly major developments are displaced by new headlines, noting that OpenAI's reported security breach involving Hugging Face had occurred just four days prior, despite feeling like a distant event from a month ago. With new releases from Opus to open-source breakthroughs dropping continuously, the hosts thank Aaron for joining the discussion before transitioning to the network outro and standard financial disclosures.
- The hosts emphasize the unprecedented tempo of AI news, where monumental events like security breaches and frontier model launches fade from memory within days. - The episode concludes with distribution notes across YouTube, Apple Podcasts, Spotify, X, and Substack, followed by standard a16z legal and advertising disclaimers.
## Notable quotes > `22:30` I generally think if you think that you've kind of eliminated the need for software engineers, there's just no chance you're being ambitious enough with your product roadmap. — Aaron Levie
> `00:00` I almost unequivocally take the stance that more open innovation and open weights innovation is good for AI broadly and the diffusion of AI. — Aaron Levie
> `10:00` I just don't think there's a scenario here where you can kind of close off completely to China and somehow you dramatically slow them down. — Aaron Levie
## Mentioned Aaron Levie, Theo Jaffee, Sofia Puccini, Box, Jensen Huang, NVIDIA, Anthropic, OpenAI, Google, DeepSeek, Moonshot AI, Dwarkesh Patel, The Dwarkesh Podcast, Bill & Melinda Gates Foundation, Claude 3.5 Opus, GPT-4, GPT-4o, Fable, Kimi k1.5, Grok 4.5, GPT-5.6, Llama, Cognition, Factory, Cursor, Replit, Hugging Face, xAI
## Video description Box co-founder and CEO Aaron Levie joins MTS hosts Theo Jaffee and Sofia Puccini to make the case for open-weight AI, unpack the economics of open versus closed models, and explain why he believes more openness could strengthen rather than undermine the U.S. AI ecosystem. Aaron argues that open models create more use cases, push closed labs to innovate faster, and don't fundamentally change where the economics of AI ultimately accrue. They debate model distillation, America's competition with China, why restricting access may simply accelerate competing AI ecosystems, and whether U.S. labs should begin releasing open-weight versions of previous-generation models. They also get into what the latest frontier models mean for knowledge work, how AI has changed software engineering at Box, and why Aaron believes companies cutting engineers may simply not be ambitious enough. Finally, they discuss why enterprises are unlikely to bet on a single model and why the layer that routes between models, data, and workflows could become increasingly valuable.
[Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/fa4d1de0-221c-4305-9b44-d8ee5ef2a0db/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=fa4d1de0-221c-4305-9b44-d8ee5ef2a0db&feed=JGE3yC0V)