CXOTalk

Companies running AI must decide which work to send to the frontier labs and which to run on open-weight models. Raffi Krikorian, Mozilla's Chief Technology Officer, explains open-weight AI for the enterprise: what open-weight and open-source models are, which work belongs on frontier models, and when running your own models pays off. The conversation covers cost, security, backup plans, token pricing, and lock-in.

YOU'LL DISCOVER
✅ Why Raffi compares an open-weight model to a compiled binary: you can download and run it, but you cannot see what went into it, unlike a fully open-source model.
✅ Where to draw the line: frontier models for scientific research, open-weight models for everyday tasks at about 35 to 50 times lower cost.
✅ How to find your own line: record and log every prompt and response, so those traces show which work a cheaper model can handle.
✅ Why openness leads to more security, not less, with Signal as the example.
✅ Why Raffi expects the moat to shift from building models to using them: fine-tuning, deployment, and the tooling around the model.
✅ How to run open models today, from NeoCloud API providers to Ollama, vLLM and llamafile on a MacBook, Mac Studio or DGX Spark, and why RAM is king.
✅ Why Raffi tells CTOs not to switch providers today, but to start evaluating a backup plan for outages, cost spikes, and lowered token limits.
✅ Why today's token prices are subsidized ahead of the labs' IPOs: a $200 subscription that one Mozilla engineer measured at about $6,000 a month in tokens.

⏱️ TIMESTAMPS

0:00 Introduction
0:14 Open weight versus open source models
5:11 Rent the frontier, own everyday AI
12:21 Openness leads to more security, not less
14:45 The moat shifts from model to usage
18:28 Open models are getting easier to run
27:42 Start evaluating a model backup plan today
32:51 When enterprises should move to open models
38:51 Subsidized token prices will not last
43:13 IP rights and inspection at scale
47:00 Open-weight competition, China, and safety
51:08 Token price hikes and context lock-in

CXOTalk streams live with audience questions. Join us live and ask the guests your own questions: https://www.cxotalk.com/subscribe

Show notes, transcript, and summary: https://www.cxotalk.com/episode/mozilla-cto-open-source-and-open-weight-ai-rent-or-own

Episode 933 | Recorded October 9, 2026

#CXOTalk #OpenWeightAI #OpenSourceAI #EnterpriseAI #AIStrategy #Mozilla #OpenModels #AICosts #CTO #AIInfrastructure

AI agents now act on more decisions than any manager can review, so the choice of which work goes to an agent and where a person stays in charge decides the result. Paula Goldman, Salesforce's Chief Ethical and Humane Use Officer and EVP of Product, explains why the human in the loop model fails for agents and what replaces it: human at the helm. Drawing on her book, Manage the Machine, she walks through choosing use cases, setting an agent's boundaries, audit trails, token spend, and the accountability structure that CEOs must put in place.

YOU'LL DISCOVER

✅ The test that decides a use case: not whether AI can do the work but whether it should, with the healthcare scheduling and 1-800Accountant tax examples
✅ Why the human in the loop idea from Cold War missile defense breaks when one agent run leaves 50,000 data points no person can verify one by one, and what human at the helm means instead
✅ Why a public tracker that now counts well past 1,000 court cases of AI-hallucinated citations points to overtrust, and the junior employee method for expanding an agent's scope
✅ The first step in setting up an agent: its goal, off-limits topics, escalation triggers, and domain tripwires such as brand voice
✅ How deterministic controls and probabilistic models work together, and why audit trails give you receipts for what an agent did
✅ The IKEA couch story and the human goes first method that stops AI from reverting to the mean
✅ How to manage token spend by allocating it per use case and matching the model to the job
✅ Advice for CEOs: a real accountability structure for AI and employees at the table

⏱️ TIMESTAMPS

0:00 Introduction
0:27 Not just can AI, but should it
6:57 Human at the helm, not the loop
11:26 AI has data, humans have context
16:24 AI finds insights, people catch mistakes
19:45 Demand AI you can monitor
24:16 Set the agent's boundaries in advance
28:58 Audit trails give you the receipts
34:58 Human goes first, then the AI
40:35 AI is simpler than it seems
45:12 Society sets the floor, companies experiment
48:12 AI needs a clear accountability structure

Join us live and ask the guests your own questions: https://www.cxotalk.com/subscribe

Show notes, transcript, and summary: https://www.cxotalk.com/episode/salesforces-chief-ethical-and-humane-use-officer-how-to-build-great-ai-use-cases

Episode 931 | Recorded September 11, 2026

#CXOTalk #AIAgents #AgenticAI #EnterpriseAI #Salesforce #AIEthics #AIGovernance #ResponsibleAI #CIO

Patrick Bangert, Chief AI Officer at Occidental Petroleum, advises enterprise leaders to avoid running AI pilots altogether. He explains that most enterprise AI initiatives fail during the transition from a successful pilot to a full-scale deployment. Instead, he recommends securing funding and management support from the outset and viewing adoption as the key to converting projected returns into actual benefits. His team manages hundreds of AI systems across Oxy's physical operations, including CO2 injection in the Permian and the deployment of cameras and edge computers in the field. The discussion covers how these systems are trained, secured, funded, adopted, and evaluated.

YOU'LL DISCOVER

✅ Why POC should stand for proof of competency, not proof of concept, and why Patrick can validate a concept in ten minutes by searching for it
✅ How free or discounted pilots leave client management hearing about a project for the first time at the results meeting, and what that does to the budget cycle
✅ The change management method that gets field operators to use a tool: stakeholders involved before the project starts, two-week sprints, co-creation and training
✅ Why revenue impact comes first, cost savings second, and adoption is the metric that lets either one happen
✅ How machine learning and optimization set CO2 injection across dozens of wells 24/7 while LLMs only handle the communication layer
✅ Why cleansing the data is the secret sauce of training, with physics rules and hard engineering limits as guardrails
✅ Why Patrick errs on the side of buying infrastructure software rather than building it, and why Oxy does much of its LLM work on open weight models with no token cost
✅ Patrick's rule that if you cannot measure it, it is zero, and the two pieces of hype leaders should stop believing

⏱️ TIMESTAMPS

0:00 Introduction
0:10 LLMs communicate; they do not decide
2:53 Field conditions push computing to the edge
5:05 Cleaning the data is the secret sauce
11:09 Buy rather than build; run open weights
17:58 Proprietary models win; vendors need scientific proof
21:25 AI results still need human validation
26:45 Scaling AI is a hardware problem
30:14 Adoption is how the revenue KPI happens
34:02 Skip the pilot, commit on day one
39:41 If you cannot measure it, it's zero
42:07 AI fails for human reasons, not technology

Join us live and ask the guests your own questions: https://www.cxotalk.com/subscribe

Subscribe for weekly conversations with leading business and technology leaders.

Get the CXOTalk newsletter: https://newsletter.cxotalk.com

Show notes, transcript, and summary: https://www.cxotalk.com/episode/occidental-petroleums-chief-ai-officer-ai-at-scale-in-the-physical-world

Episode 932 | Recorded September 18, 2026

#CXOTalk #EnterpriseAI #AIAdoption #PilotPurgatory #IndustrialAI #OilAndGas #ChiefAIOfficer #AIStrategy #CIO

Check out Gartner Symposium: https://cxo.news/KGg8XY
===
Tamar Yehoshua is Chief Product and AI Officer at Atlassian, where AI agents are now direct users of the software her team builds. She led product at Google Search and Slack before joining Atlassian, and in this conversation she explains how the company designs products for both people and agents: the teamwork graph that gives an agent the organizational context it arrives without, who stays accountable when a chain of agents goes wrong, how Atlassian measures AI spending with its CFO, and why she never believed agents would kill SaaS.

YOU'LL DISCOVER

✅ Why an AI agent arrives like someone joining your organization with no background whatsoever, and how the teamwork graph (over 200 billion entities) supplies the missing context
✅ The result of Atlassian's A/B tests: 44% higher quality and 48% lower token usage when coding agents use the teamwork graph, versus MCP alone
✅ How Mercedes-Benz built a Rovo agent that cut manual bug triage work for engineers by 85%
✅ The 10x question Atlassian asks before automating a process, instead of settling for a 10 to 20% incremental gain
✅ How Atlassian measures AI value with the CFO: PRs deployed per engineer, features developed, and whether a growing token budget delivers more value
✅ Why vibe-coded replacements for enterprise SaaS "just don't last" at large companies: support, security, and enterprise readiness
✅ The accountability principle for agent chains: the human who builds the first agent stays responsible, with agents increasingly acting as guardrails for other agents
✅ Why Atlassian's AI gateway routes each job to the right model on quality, latency, and cost instead of locking into one model

⏱️ TIMESTAMPS

0:00 Introduction
0:26 Building products for humans and agents
5:50 The teamwork graph and agent context
13:33 Human oversight and agent guardrails
18:26 Moving fast versus getting it right
21:56 Efficiency first, then new capabilities
26:25 Tokens, vibe coding, and SaaS survival
33:00 Agent accountability and the software lifecycle
38:53 Simpler interfaces and visualizing agent work
41:49 How AI reshapes leadership and hierarchy
48:07 Reimagining processes and the Forge ecosystem
50:44 Advice for CIOs and managing AI costs

CXOTalk streams live with audience questions. Join us live and ask the guests your own questions: https://www.cxotalk.com/subscribe

Subscribe to CXOTalk for weekly conversations with leading business and technology leaders.

Get the CXOTalk newsletter: https://newsletter.cxotalk.com

Show notes, transcript, and summary: https://www.cxotalk.com/episode/atlassian-chief-ai-officer-building-products-for-agents-and-humans

Episode 930 | Recorded August 28, 2026

#CXOTalk #AIAgents #Atlassian #EnterpriseAI #ArtificialIntelligence #AgenticAI #ChiefAIOfficer #ProductManagement #DigitalTransformation #SaaS

Check out Gartner Symposium: https://cxo.news/KGg8XY
========
More than 90% of the code at Snap is now written by AI agents, and a code review agent checks most of it within ten minutes. Saral Jain, SVP, Head of Engineering and CIO at Snap Inc., built the system that makes that safe for nearly a billion monthly users. He explains the golden path of 14 managed agents, the guardrails that account for more than half of the investment, and why accountability still rests with a named human for every change that ships.

YOU'LL DISCOVER

✅ Why Snap funds 14 blessed agents on a golden path instead of letting every team build its own
✅ How Casper turns a Slack conversation into a sandboxed pull request that CodePal reviews before a human sees it
✅ Why more than half of Snap's agent investment goes into evals, code review and rollback rather than into building
✅ The risk-tiering test that separates an internal dashboard from a privacy-sensitive change
✅ Why Saral says the risk is not AI-generated code, it is unowned code
✅ How Snap measures value on outputs, with commits up 75% year over year and severe outages down 57%
✅ Why context beats model 100 out of 100 times, and what context actually means inside a 15-year-old codebase
✅ Four pieces of advice for CIOs, ending with why adoption has to start at the top

⏱️ TIMESTAMPS

0:00 Agents write the code, engineers own it
7:17 Agents as teammates with defined personas
12:52 Systems for agents, features for humans
16:09 How Snap governs thousands of agents
22:39 Invest in guardrails before building agents
29:46 Output metrics and the coordination tax
33:08 Context you can give, judgment you cannot
38:14 The risk is unowned code
44:58 Working with unpredictable AI systems
47:13 Agents beyond engineering, and human skills
50:24 Token spend and model routing
52:35 Advice for leaders and the risks ahead

Subscribe for weekly conversations with leading business and technology leaders.

Get the CXOTalk newsletter: https://newsletter.cxotalk.com

Show notes, transcript, and summary: https://www.cxotalk.com/episode/snap-agentic-ai-software-development-at-scale

Episode 929 | Recorded August 21, 2026

#CXOTalk #AgenticAI #SoftwareDevelopment #EngineeringLeadership #AICodeReview #Snap #CIO #AIGovernance #DeveloperProductivity #snapchat