Agent Reliability and Salesforce Concerns

On Friday’s show, the DAS crew discussed what real AI productivity looks like in 2025, where agents still break down, and how the biggest platforms are pushing assistants into products people already use. They covered fresh survey data on AI at work, Salesforce’s push for more deterministic agents, OpenAI’s role based prompt packs, a reported Waymo in car Gemini assistant, Meta’s non generative “world model” work, holiday AI features, and the ongoing Lovable vs Replit debate for building software fast. The episode also touched on AI infrastructure and power constraints, plus how teams should think about curriculum, playbooks, and repeatable workflows in an AI first world.

Key Points Discussed

Lenny Rachitsky shared survey results from 1,750 tech workers on how AI is actually used at work

55 percent said AI exceeded expectations, 70 percent said it improves work quality

More than half said AI saves at least half a day per week, founders reported the biggest time savings

Designers reported the weakest ROI, founders reported the strongest ROI

92.4 percent reported at least one significant downside, including reliability issues and instruction following problems

Salesforce leaders highlighted agent unreliability and “drift”, AgentForce is adding more deterministic rule based structures to constrain agent behavior

OpenAI Academy published prompt packs grouped by job role, showing how OpenAI frames “default” use cases

Waymo is reportedly working on a Gemini powered ride assistant, surfaced via a discovered system prompt in app code

Meta’s VLJEPA work came up as an example of non generative vision models aimed at world understanding, not image generation

The crew debated Lovable and Replit as fast paths from idea to working app, including where each still breaks down

Timestamps and Topics

00:00:17 👋 Opening, Boxing Day, setting up the “is AI delivering ROI” question

00:02:20 📊 Lenny Rachitsky survey, who was sampled, what it measures

00:05:44 ✅ Top findings, time saved, quality gains, ROI split by role

00:07:33 🧩 Agents and reliability, Salesforce view on drift, AgentForce guardrails

00:10:25 🧰 OpenAI Academy prompt packs by role, why it matters

00:12:07 🚗 Waymo and a Gemini powered ride assistant, system prompt discovery

00:13:05 👁️ Meta VLJEPA, non generative vision and “world model” direction

00:15:47 🎄 Holiday AI features, Santa themed voice and image moments

00:16:34 ⚡ Power and infrastructure constraints, wind and solar angle for AI buildout

00:20:05 🛠️ Lovable vs Replit, speed to product and practical tradeoffs

00:25:00 💻 Claude workflow talk and migration friction (real world setup issues)

00:30:00 ☁️ Cloud strategy, longer prompts, and getting useful outputs from big context

00:38:00 🎓 Curriculum and workforce readiness, what to teach and what to automate

00:40:10 📚 Wikipedia, automation patterns, and reusable knowledge sources

00:43:10 📓 Playbooks and repeatable processes, turning AI into a system not a novelty

00:51:40 🏁 Closing and weekend sendoff