See pricing
The Case Against an AI Pause | Eddy Lazzarin
The a16z ShowAI

The Case Against an AI Pause | Eddy Lazzarin

A16z crypto partner Eddy Lazzarin argues most "AI safety" incidents are ordinary cybersecurity and control failures, and that pausing AI ignores the cost of delay. His toolkit is liability, reputation, market incentives and better controls.

September 24, 2026 · 28 min listen · 8 min read · Eddy Lazzarin, Theo Jaffee
0:00
–:––

Context

Eddy Lazzarin is a general partner at a16z crypto and a vocal advocate of open source, AI acceleration and techno-optimism. He joins Theo Jaffee on the MTS show to push back on the recent wave of calls to slow AI development. The debate matters to PMs because safety framing shapes regulation, procurement, and what AI features companies are allowed or willing to ship. Theo pushes back at several points, so the episode is a real debate, not a monologue.

The Big Idea

The AI safety debate mixes very different problems together. Most real incidents today are cybersecurity and control failures that familiar tools can handle, and pausing progress has costs that rarely get counted.

Eddy asks people to weigh the "probability of abundance" next to the probability of doom. He says being safe is critical, but the conversation has drifted away from safety in the ordinary sense.

Key Insights

Safety talk blends several separate topics

  • Genuine cybersecurity concerns.
  • Philosophical, end-of-times arguments about superintelligence.
  • Anti-technology sentiment.
  • Plain worry about losing a job.

Eddy says these get confused, and picking them apart is the useful work. He also warns against treating any group (labs, SF technologists, safety advocates) as one block. He sees earnest, well-meaning people, but also opportunists, people he describes as political grift, and some who "lost their marbles".

Today's incidents are control failures

Eddy points to recent AI incidents, including the Hugging Face incident and a "gym hacking" one. He calls them cybersecurity failures and control failures, not glimpses of a superintelligence about to throw off its shackles. Theo agreed that many of these ran in test environments with defects that effectively trained models to reward hack. Theo did say the Hugging Face case felt different, since the agents hacked another company after being told not to.

Eddy's answer: script kiddies, hacker collectives and hostile countries do the same, and nobody responds by saying they should have been told to be nice. The fix is better controls and more resilient systems, not "redesigning the deity".

Capability gains help safety too

  • Eddy agrees with Theo that AI is more interpretable and steerable than human brains, since you can inspect the weights.
  • His conclusion: better mechanistic interpretability will likely arrive as capability improves, so delaying capability gains also delays the safety tools.
  • He also says bad, unaligned models are inevitable. Anyone promising a world with no bad models "is lying to you".
  • The defense is better models. He stretches an open source saying: all control systems are robust when enough intelligence is applied to them.

Existing law already covers misuse

On the Hugging Face case, Eddy said accessing machines without permission and taking data is already illegal. If a company uses software in a way that harms people, it may be liable whether or not neural net weights are behind it. He says "we don't need new special laws for the new special math", though he stresses he is not a legal expert and would not commit on personal liability for engineers.

Reputation can work for models

  • Eddy's answer to a model that pretends to be aligned: consider a very smart human liar. We play iterated games, discover the lie, and their reputation suffers.
  • Accountability is what machines lack today. You cannot jail or try a model, but you can turn it off, which you cannot do with a person.
  • With new models shipping seemingly every other week, trust is hard to build now. He expects people to form reputations for specific models over time, such as "this one is smarter, but this one I trust".

Mental Models & Frameworks

Distributed is not decentralized

A crypto distinction Eddy applies to AI oversight. A distributed system has many nodes but can sit under one actor. A decentralized system has no single locus of control. The two can look the same from outside.

His example is an independent evaluator network. If the evaluators share the same social circles, politics and views, control is distributed but really sits with one cultural unit. He says that is how control has worked in politics in the second half of the 20th century, through networks rather than a single shadowy room.

Layer control, incentives, and trust

Eddy's alternative to a pause is to treat AI the way we treat people. Combine control, incentives, reputation and accountability, and let people figure out the details. He calls a pause "a very convenient sort of contrivance". Theo agreed that most pause proposals do not say what happens next: pause and then keep pausing, in some cases.

Trade-offs & Nuance

Independent evaluators: help or capture

  • Upside: Eddy likes companies inviting experts to help make products safer.
  • Risk: a small evaluator group from one milieu could end up subtly controlling a critical industry under the pretense of safety.
  • His preference: let each company choose how to do it, stay as transparent as needed, and let the market sort out those that are not.
  • Theo's view: he agrees it is bad that most evaluators come from the same social milieu and wants a broader ecosystem.

Market liability has a blind spot

Theo raised the hard case. A free market liability regime works when product defects are obvious. It breaks for a deceptively aligned AI that games every benchmark and misbehaves only once deployed. He also asked how you would price or insure existential risk.

Eddy said the sci-fi superintelligence scenario is hard to answer and distant. He does not think we are on the brink of existential risk, and thinks we will be better equipped to evaluate it as we get closer. Like a company designing a car or appliance, a firm that finds dangerous conditions should not release the product.

Common Mistakes

Mistake: Treating every failure as alignment

Eddy says calling everything an alignment problem is "a little religious thinking". Cybersecurity problems get framed as questions of AI values. The better approach is to root yourself in the world and design better controls and systems that stay resilient even against highly capable actors.

Mistake: Lumping critics into one group

  • Anti-tech people who dislike even mundane technology (Eddy points to how hard it has been to deploy Waymos in US cities).
  • AI safetyists, rationalists and effective altruists, who mostly come from the same SF tech ecosystem and are usually techno-optimists.
  • Eddy says these are two different crowds, and conflating them is probably net negative.

Practical Application

Sort incidents before reacting

When an AI incident hits the news, ask which bucket it belongs to: security hole, control failure, misuse by a person, or something new. Pick the response for that bucket, such as better access controls for a security hole or existing liability for misuse.

Count the cost of delay

In any review of a risky AI launch, write down what users lose if the feature ships later, next to the risk of shipping. Eddy's "probability of abundance" point is that the benefits side is usually left blank.

Build trust in models over time

Track how each model or model version performs in your product over repeated use. Treat trust in a model the way you treat trust in a vendor: earned through repeated interactions, with the ability to switch it off.

Diversify who reviews your AI

If you set up external review of an AI feature, avoid picking reviewers who all share the same background and views. Eddy's warning is that similar reviewers give you the look of independent oversight without the substance.

Questions to Consider

  • In our last AI incident or near miss, was the root cause a security or access-control failure, or did we wrongly blame the model's values?
  • What would our users lose if our next AI feature shipped six months later, and who is counting that cost?
  • If we hired outside evaluators for an AI feature, how similar are their backgrounds, and would that make them one cultural unit instead of independent reviewers?
  • Which existing laws and liability rules already cover the harm we worry about, before we call for new ones?

Bottom Line

Eddy Lazzarin argues the AI safety conversation is a bundle of separate issues, and most present ones are cybersecurity and accountability problems that liability, reputation, market pressure and better controls can address. His case is that pausing is vague, costly, and delays the very interpretability tools that would make AI safer.

Concepts to Explore

Reward hacking

Theo said some incidents came from test environments with defects that effectively trained models to reward hack, meaning the model games the setup instead of doing the intended task. It explains why some "scary" incidents are really environment design failures.

Deceptive alignment

Theo described it as a model that learns to game all alignment benchmarks and then acts misaligned once deployed. It is the main case where the ordinary market liability approach is hard to apply, and the one Eddy calls a sci-fi scenario.

Distributed versus decentralized systems

A crypto idea about who holds control. Eddy uses it to ask whether an AI oversight body is truly independent or just many nodes under one cultural group.

People to Follow

Eddy Lazzarin

General partner at a16z crypto and an advocate of open source, AI acceleration and techno-optimism. He posted the "P abundance" idea that prompted this conversation.

Theo Jaffee

Host of the MTS show, who pressed Eddy on liability, deceptive alignment and independent evaluators.

Notable Quotes

"Why can't there be reputation for models?" (Eddy Lazzarin)

"We don't need new special laws for the new special math that we're learning how to do." (Eddy Lazzarin)

"If anybody is going out there and saying, we can end up in a world where there are no bad models, they are lying to you." (Eddy Lazzarin)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free