Should AI oversight work through government pre-approval of model releases or through outcome-based liability on the companies deploying them?

Tomasz Tunguz · Apr 21, 2023

In AI, the moat is not the data moat — it is better execution, and startups can beat incumbents despite their distribution advantage

Whenever machine learning comes up people ask about the data moat, but the answer is what it has always been: build a better product and get it to market, as Notion did in documents and Snowflake did in databases against big incumbents

Scope: revised from his earlier view that incumbents would win the whole thing

37:58 20VC: Who Wins in AI; Startup vs Incumbent, Infrastructure vs Application Layer, Bundled vs Unbundled Providers | From 150 LP Meetings to Closing $230M for Fund I; The Fundraising Process, What Worked, What Didn't and Lessons Learned with Tomasz Tunguz

Matt Clifford · Jul 1, 2024

The EU AI Act is a mistake because it tries to anticipate the future, labels a set of things high risk, and creates an enormous burden on companies innovating in those areas

The legislation was written by imagining everything that could go wrong and prohibiting it in advance

29:43 20VC: LLMs Are Reaching a Stage of Diminishing Returns: What is the Next S Curve | The Bull & Bear Case for China's Ability to Challenge the US' AI Capabilities | How AI Changes the Future of War & How Agents Will Reshape Society with Matt Clifford @ EF

Ethan Mollick · Jul 31, 2024 · hedged

The right approach to AI regulation is fast follow-up regulation rather than pre-regulation: don't try to regulate in advance, but watch closely and put rules in place quickly as problems emerge

With a new technology you don't know what the problems will be or what it's good and bad at, so you have to learn from deployment

Scope: endorses Joshua Gans's model as 'probably right'; acknowledges government isn't built to react fast or cooperate well with industry

19:34 20VC: Is More Compute the Answer to Model Performance | Why OpenAI Abandons Products, The Biggest Opportunities They Have Not Taken & Analysing Their Race for AGI | What Companies, AI Labs and Startups Get Wrong About AI with Ethan Mollick

Ethan Mollick · Jul 31, 2024

Regulating in advance to stop harms that haven't occurred is the wrong approach, because current models clearly cannot cause them — a runaway superintelligence will not emerge from Llama 3.1 — but zero regulation and no scrutiny is equally untenable; a balance is needed

You want the technology developed in democratic societies and used in democratic ways, so you want continued growth, but it's weird to pretend there is no downside risk

Scope: describes himself as a technology optimist; holds both sides: pre-regulation wrong, no-scrutiny also wrong

21:08 20VC: Is More Compute the Answer to Model Performance | Why OpenAI Abandons Products, The Biggest Opportunities They Have Not Taken & Analysing Their Race for AGI | What Companies, AI Labs and Startups Get Wrong About AI with Ethan Mollick

Joelle Pineau · Nov 3, 2025

Regulation should not be expected to run ahead of the technology; the technology needs creative space first, with guardrails developed from real learnings

AI is an incredibly young and fast-moving field and governments need to benefit from practitioners' knowledge to make good policy

24:51 20VC: Cohere's Chief AI Officer on Why Scaling Laws Will Continue | Whether You Can Buy Success in AI with Talent Acquisitions | The Future of Synthetic Data & What It Means for Models | Why AI Coding is Akin to Image Generation in 2015 with Joelle Pineau

Eiso Kant · Oct 7, 2024

AI regulation should target end-user applications and hold companies accountable for how their technology affects users, rather than limiting how much compute a model can be trained on

Harm comes from use, not from the underlying technology — 'it's not the database that does harm, it's how it's used' — and this is how every previous technology has been regulated

Scope: believes the world is already moving in this direction

60:12 20VC: Raising $500M To Compete in the Race for AGI | Will Scaling Laws Continue: Is Access to Compute Everything | Will Nvidia Continue To Dominate | The Biggest Bottlenecks in the Race for AGI with Eiso Kant, CTO @ Poolside

Anastasios Angelopoulos · Aug 3, 2026

Requiring a central government body to approve each model release is a crazy idea; the better approach is to keep the strongest scientists in private companies and use rules and large liability for outcomes like corporate data leaks to incentivize good safeguards

Government agencies are bureaucratic and ill-suited to this; the capitalist incentive system does this kind of work well

31:40 20VC: 70% of Neolabs Will Die | There Will be a $100BN US Open-Source Model | Data is a Trillion $ Market | Governments Cannot Regulate Models: It is Too Late | The Cyber Attacks to Come Will be Insane with Anastasios Angelopoulos @ Arena

Anastasios Angelopoulos · Aug 3, 2026

AI oversight should be outcome-based regulation of businesses — huge fines and scrutiny when a company's AI causes a breach — rather than a government process charged with preventing incidents, because government is not technically capable of that

A neutral body is understandable in principle but hard to make work outside a company; you want the incentive system to sort itself out

Scope: acknowledges the appeal of a neutral body

32:25 20VC: 70% of Neolabs Will Die | There Will be a $100BN US Open-Source Model | Data is a Trillion $ Market | Governments Cannot Regulate Models: It is Too Late | The Cyber Attacks to Come Will be Insane with Anastasios Angelopoulos @ Arena

Joelle Pineau · Nov 3, 2025

Governments should define the standards everyone agrees on while companies build and deploy the verification solutions at scale

24:22 Government defines standards companies build the verification

Nick Frosst · Sep 1, 2025

The worst thing AI regulation could do is, out of a mistaken belief that large language models are digital gods, pick a gameable benchmark as a proxy for AGI and shut down development on that basis

Large language models are not digital gods, and benchmarks can be trained to do much better or worse, so fixating on them tells you nothing about how the technology can be used or misused

58:51 Benchmark based agi proxies are gameable and should not gate regulation

Demis Hassabis · Apr 7, 2026

Government must ultimately be the verifying authority for AI systems, with technical auditing done by AI safety institutes, and leading research countries should each stand up an equivalent well-staffed body

Independent, technically capable bodies staffed with high-quality researchers are needed to evaluate and audit models against benchmarks

19:39 Government verification via technically staffed safety institutes