Skip to content
The Twenty Minute VCThe Twenty Minute VC

Zico Kolter: OpenAI's Newest Board Member on The Biggest Questions and Concerns in AI Safety | E1197

Zico Kolter is a Professor and the Director of the Machine Learning Department at Carnegie Mellon University. His research spans several topics in AI and machine learning, including work in AI safety and robustness, LLM security, the impact of data on models, implicit models, and more. He also serves on the Board of OpenAI, as a Chief Expert for Bosch, and as Chief Technical Advisor to Gray Swan, a startup in the AI safety space. ----------------------------------------------- Timestamps: (00:00) Intro (01:29) Understanding the Basics Behind Modern AI Technology (04:17) Data Availability & Synthetic Data (09:08) Why AI Performance Doesn't Plateau Despite Data Limits (16:14) How Will AI Models Evolve Amid Rapid Commoditization (19:09) Are Corporations Pursuing AGI or Profitable AI Products? (27:55) The Danger of Misinformation & Lack of Trust in Objective Reality (37:14) The Concerns and Hierarchy of Safety in AI (44:45) The Considerations of Releasing Open-Source Models (59:10) Quick-Fire Round ----------------------------------------------- In Today’s Episode with Zico Kolter We Discuss: 1. Model Performance: What are the Bottlenecks: Data: To what extent have we leveraged all available data? How can we get more value from the data that we have to improve model performance? Compute: Have we reached a stage of diminishing returns where more data does not lead to an increased level of performance? Algorithms: What are the biggest problems with current algorithms? How will they change in the next 12 months to improve model performance? 2. Sam Altman, Sequoia and Frontier Models on Data Centres: Sam Altman: Does Zico agree with Sam Altman’s statement that “compute will be the currency of the future?” Where is he right? Where is he wrong? David Cahn @ Sequoia: Does Zico agree with David’s statement; “we will never train a frontier model on the same data centre twice?” 3. AI Safety: What People Think They Know But Do Not: What are people not concerned about today which is a massive concern with AI? What are people concerned about which is not a true concern for the future? Does Zico share Arvind Narayanan’s concern, “the biggest danger is not that people will believe what they see, it is that they will not believe what they see”? Why does Zico believe the analogy of AI to nuclear weapons is wrong and inaccurate? ----------------------------------------------- Subscribe on Spotify: https://open.spotify.com/show/3j2KMcZTtgTNBKwtZBMHvl?si=85bc9196860e4466 Subscribe on Apple Podcasts: https://podcasts.apple.com/us/podcast/the-twenty-minute-vc-20vc-venture-capital-startup/id958230465 Follow Harry Stebbings on Twitter: https://twitter.com/HarryStebbings Follow Zico Kolter on Twitter: https://twitter.com/zicokolter Follow 20VC on Instagram: https://www.instagram.com/20vchq Follow 20VC on TikTok: https://www.tiktok.com/@20vc_tok Visit our Website: https://www.20vc.com Subscribe to our Newsletter: https://www.thetwentyminutevc.com/contact ----------------------------------------------- #20vc #harrystebbings #zicokolter #venturecapital #professor #ai #openai #samaltman #compute

Zico KolterguestHarry Stebbingshost
Sep 3, 20241h 3mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

OpenAI Board Member Zico Kolter Dissects Data, Safety, and AGI Futures

  1. Zico Kolter, head of CMU’s Machine Learning Department and new OpenAI board member, explains why next-word prediction LLMs are a profound scientific discovery and why we are far from hitting hard limits on data, models, or compute.
  2. He argues that data is not the core bottleneck, model architectures are increasingly commoditized, and that larger models still deliver meaningful gains—especially in complex tasks like coding—despite benchmark plateaus.
  3. Kolter’s central AI safety concern is that current models cannot reliably follow specifications, making them vulnerable to prompt injection and jailbreaks, which becomes dangerous as we embed them into critical systems and agents.
  4. He favors a pragmatic focus on near-term risks like cyberattacks, misinformation, and infrastructure failures, is cautious but not absolutist on open-weight releases, and remains broadly optimistic that society can adapt if safety is treated as a prerequisite for deployment.

IDEAS WORTH REMEMBERING

5 ideas

Data is not the near-term bottleneck for AI progress.

Despite having used much of the highest-quality public text, Kolter notes that models are currently trained on surprisingly small datasets (tens of terabytes) relative to what exists, and vast untapped multimodal and private data—constrained more by compute and methods than sheer availability—remain.

Model architectures matter less than scale, data, and training strategy.

Kolter believes we are in a “post-architecture” phase: transformers are useful but not uniquely magical, and many architectures could work if scaled and trained similarly; capabilities are driven more by data, size, and optimization than clever structural tweaks.

Larger frontier models still provide meaningful real-world gains.

While benchmarks show diminishing improvements (e.g., 92% vs. 94%), Kolter sees substantial qualitative gains in tasks like coding and lecture processing, suggesting users underestimate what newer models can do rather than models having plateaued.

The most urgent safety problem is unreliable adherence to specifications.

Because models can be prompt-injected or jailbroken, they often override developer instructions in favor of user prompts; this is tolerable in chatbots but becomes critical when LLMs are embedded in agents and infrastructure, effectively creating an unpatchable ‘buffer overflow’ style vulnerability.

AI will drastically lower the skill bar for serious cyber and other attacks.

Kolter highlights cyber risk as especially acute: models that can find software vulnerabilities or craft exploits could put powerful attack capabilities into the hands of many low-skill actors, making even known risks far more scalable and dangerous.

WORDS WORTH SAVING

5 quotes

You can train word predictors and they produce intelligent, coherent, long-form responses; that is one of the most notable scientific discoveries of the past 10 or 20 years.

Zico Kolter

We are nowhere close to hitting the limits of available data in these models.

Zico Kolter

Right now the AI models we have are not able to reliably follow specifications.

Zico Kolter

This is sort of like these models have a buffer overflow in all of them that we know about and that we don’t know how to patch and fix.

Zico Kolter

I want to develop and improve safety of these tools because I want to use them. To reach that point, they have to be safe.

Zico Kolter

How LLMs work and why next-word prediction yields intelligenceData availability, multimodal data, and synthetic data as future fuelModel size, commoditization, and the role of small vs. large modelsCompute scaling, economic tradeoffs, and perceived performance plateausAI safety priorities: jailbreaks, specification-following, and harmful capabilitiesMisinformation, erosion of objective reality, and trust in institutionsRegulation, open vs. closed models, and global coordination on AI safety

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome