alphalist.CTO Podcast - For CTOs and Technical Leaders

Mathias Lechner did his research in Vienna on the nervous system of C. elegans, a worm with just 302 neurons. That work became liquid neural networks, a self-driving demo that steered a car with 19 neurons, and eventually Liquid AI, the MIT spinoff he co-founded and leads as CTO. AMD led Liquid AI's $250 million round, and its models power the in-car assistant in Mercedes' new CLA.

In this episode, Tobi and Mathias dig into what it really takes to run AI without the cloud. Mathias explains why Liquid AI starts with the hardware instead of model size, why memory bandwidth and memory size matter as much as FLOPS on a phone or in a car, and why he believes there's no long-term technical moat in AI, only a race where silicon partnerships give you the edge. He also shares his own coding agent setup, how the team judges training data, and what he'd tell his 2017 self.

CTOs will come away with a clear picture of which workloads will move on-device and which stay in the cloud, what limits local inference today, and how to start testing small open-weight models themselves.

- From a worm's nervous system to liquid neural networks and Liquid AI
- Why a 19-neuron network could steer a car while the industry scaled up
- No long-term moat: silicon partnerships and access to pre-release chips
- Speed is non-negotiable: designing models around the hardware
- Memory bandwidth, memory size and FLOPS: the limits of on-device AI
- Cars, phones and the Mercedes in-car assistant
- Data quality, evals and his YOLO-mode coding agent setup

Sri Viswanath was Atlassian's first ever CTO. He joined right after the IPO and led the move from on-prem to cloud as revenue grew from $500 million to $2.5 billion and engineering scaled past 7,000 people. After a stint as a VC, he became a founder for the first time. His company, Sycamore, has raised a $65 million seed round to build an operating system for enterprise AI agents.

In this episode, Toby and Sri dig into what changes when software can be built, deployed and rewritten autonomously. Sri explains why software starts decaying the day it ships, and why some Sycamore customers are already on version 50 of their apps. He also covers why agents should earn trust over time instead of getting it on day one, and why he believes the more governance you have, the faster you can go. He also looks back on Atlassian, including the cloud migration call he still calls his worst decision.

CTOs will come away with a practical framework for moving agents into production: progressive trust, zero-trust agent identity, verification over code review, and the systems thinking that separates real transformation from automating one workflow at a time.

- Atlassian's cloud transformation and why it needed a CTO after 14 years
- Why software decays after deployment and how a closed loop reverses that
- Progressive trust: why agents are never autonomous on day one
- Governance as speed: sandboxes, five-second tokens and guarding probabilistic systems
- 7,000 engineers for an issue tracker? What scale really takes
- Systems thinking, forward deployed engineers and deciding what to build
- The engineering org of 2028 and the human-to-agent ratio

Raffi Krikorian has built teams at scale: platform engineering at Twitter, Uber's self-driving center, and the Democratic National Committee, where he was its first CTO. Today he's CTO of Mozilla, working to make AI as open as the web, and on the side he runs three rotating Claude Max subscriptions and a GPU in his closet.

Tobi and Raffi look at what agentic coding has done to engineering teams. Writing code used to be a social contract: coordinating with other teams, negotiating requirements, sharing work. Now one person with agents does all of it alone. Raffi argues that Git, built to commit bytes, doesn't fit a world where nobody types the bytes anymore, and that the tooling, processes and roles for working as a team still need to be invented.

The conversation then moves to Mozilla's strategy: building an open source AI ecosystem without owning it, convincing CTOs and CEOs (not just developers) that open-weight models are safe to deploy, and pushing for standards like portable memory so switching providers stays possible. CTOs will come away with a clear argument for building leverage into their AI stack before their teams get locked into a single vendor's playbook.

- Why Git may be the wrong tool for agentic teams
- Why we may need more engineering and product managers, not fewer
- Raffi's personal setup: a local Hermes agent, rotating subscriptions and a sandboxed Claude Code
- Mozilla's plan for an open source AI ecosystem and open standards
- Memory as the new lock-in, and a portable memory standard
- Why chatbots should be measured by usefulness, not engagement
- What Raffi would tell his 2009 self about shutting down the Twitter API

Chris Wysopal was one of the first hackers to go public. As "Weld Pond" at the L0pht hacker collective in Boston, he testified before the US Senate in 1998, where the group delivered the soundbite that they could take down the internet in 30 minutes. He also wrote the Windows version of Netcat. In 2006 he co-founded Veracode, which by his account has now analysed trillions of lines of code. Today he's the company's Chief Security Evangelist.

Tobi and Chris talk about what AI changes for attackers and defenders. Attacks are getting cheaper and faster, and a custom exploit no longer tells you a nation-state is behind it. Chris argues this follows a familiar cycle: attackers adopt a new class of tool first, defenders catch up, and parity returns. That only holds if defenders actually adopt the tools, especially underfunded organisations like hospitals, schools and utilities. They also cover the new attack surface created by agents, plugins and MCPs, why prompt injection may never be fully solvable, and how Chris would handle security debt when acquiring a small SaaS company.

CTOs will leave with a concrete shortlist: avoid memory-unsafe languages, put a package firewall in front of open source, run AI-assisted static analysis with real architectural context, and give coding agents explicit security intent rather than hoping "make it secure" does the job.

- From the L0pht and BBS culture to the professional security industry
- The BGP flaw behind "30 minutes to take down the internet"
- How AI changes the cost and speed of attacks
- Prompt injection, agents, MCPs and least privilege
- Security debt in B2B SaaS acquisitions
- Secure on first write: security intent, context and pre-merge testing
- How engineering and security roles change over the next two years

Charles Guillemet has spent nine years building Ledger's security organization from scratch — first as the founder of The Donjon, Ledger's in-house offensive security research team, then as CTO overseeing security for a company safeguarding roughly 20% of the world's crypto (per Tobi's intro framing).

The conversation centers on one uncomfortable observation: with the latest generation of frontier AI models, Ledger's researchers started finding product vulnerabilities that fifteen dedicated experts hadn't found in nine years of trying. Charles explains why this matters security has always relied on an economic asymmetry between what it costs to attack a system and what an attacker stands to gain, and AI is collapsing that asymmetry by driving the cost of finding a vulnerability toward zero.

From there, Tobi and Charles dig into what actually holds up: security-by-design, treating security as a key-management problem, zero trust, and for teams rebuilding their SDLC around AI — Charles's recommendation of small, AI-native pods spending most of their time on "harness engineering" rather than writing code. They also cover why he believes the CISO function should stay organizationally independent, even though it reports to him at Ledger today.