From Autocomplete to Agency: Novel AI Discoveries, the Mythos/Capybara Leak and What Governed Enterprise AI Really Takes

Full Transcript

Nichol Goldstein: Hello everybody. Yay for a brand new opening theme song. For those of you who did not join us last episode, nice to meet you. I am Nichol Goldstein, the Progress Community Manager.

Hinal Patel: I am Hinal. I am the senior product marketing specialist at Progress. I work with Phil and Nichol on the daily and we are really excited to share our thoughts with you.

Phil Miller: I am Phil. My real job is product marketing director. My pretend job is AI Strategist. I get to speak to customers, partners, industry peers, and recently regulators, about what is happening in the AI market and what proactive steps you can take to operationalize AI inside your business.

Nichol Goldstein: We had a community question from Jochen Bakkers. Industry breakthroughs aside, where do we go from autocomplete on steroids to actually learning and intelligence? What is the next major direction businesses will have to deal with? Swarm intelligence? A shift to small specialist models? Domain experts that upset data lakes?

Phil Miller: The cop-out answer is all of the above. But the real story is that AI has gone beyond the autocomplete stage. It is being used for novel discoveries in multiple fields. A very old geometry problem was solved with AI assistance recently. AlphaFold-class systems have made novel discoveries in genetics. We are seeing emergent intelligence inside these models. They do not have agency. The human being is still the agent of change. Whether it is a prompt or an instructional rule that we give an agent, the human is the one setting the action.

Phil Miller: There was a leak this month from Anthropic about two new models, Mythos and Capybara. They are described as a step change in coding, cybersecurity, and academic reasoning compared to current flagship models. People are already building applications with these tools that we could not build before. This week I built two business applications in minutes. They are not hardened. They are not scalable. But they show what is now possible when subject matter experts can build their own software, instead of waiting for a development team.

Editor's note: Leaked product names are flagged as unverified until Anthropic publishes official guidance.

Phil Miller: If you are using a free tier, you are roughly a year or more behind the frontier. If you pay around twenty dollars a month, you are still six to twelve months behind. Even on the top tier you sit a quarter or two behind the labs themselves. The labs are running on what is coming next. Plan your AI strategy assuming you are always catching up, not assuming the model you have today is the model you will have in six months.

Phil Miller: For fifty years, software has been deterministic. The same inputs produce the same outputs. Language models do not work that way. They are non-deterministic. They choose what to do and when. Without guardrails, one bad action becomes ten thousand bad actions because the system scales programmatically. Governance has to be designed in at the planning stage. Not after deployment.

Hinal Patel: One technique I rely on is setting a system prompt that tells the model to challenge me, find gaps in my reasoning, and refuse to validate weak prompts. The personal preferences pane in Claude makes that easy to install once and reuse across every conversation.

Nichol Goldstein: I do the same thing on the community side. I ask the model to push back if I am making an unsupported claim. That single instruction changes the entire shape of the output.

Phil Miller: Real governance is layered. It is not one prompt. It is the system prompt that sets behavior, the MCP connectors that decide which tools and data the agent can reach, role-based access on the data layer, encryption, audit logging, and infrastructure. The Progress^® Data Platform sits at the data and context layer, which is where most enterprises are weakest. The platform decides what the AI is allowed to see and act on. The model never sees what it is not allowed to see.

Phil Miller: AI is both a mirror and a lens to the human condition. It is the bias we built into the data. There is a real case where a wrongful identification by facial recognition put an innocent person in jail for months. They lost their home. The AI did not fail in isolation. A human verification step was missing. Whose fault is it when AI makes a mistake? A human's. Every time. The developer who built it, the operator who prompted it, or the organization that did not put a check in place.

Phil Miller: The missing skill is not prompt engineering. It is critical thinking. People take AI output and treat it as fact. The training every employee needs is how to ask the next question, how to verify, and how to recognize when the model has produced something that looks confident and is wrong. Investing in AI literacy is the cheapest investment a leader can make right now. It compounds across every workflow.

Phil Miller: Most brand voice problems with AI are really brand value specificity problems. If your values are not precise enough to onboard a new human employee, the AI cannot speak them either. The fix is to translate templates, approved language, and style markers into machine-readable form. We have a hotel chatbot example where the bot was friendly, but off-brand, because the brand never wrote down what on-brand means in operational detail.

Phil Miller: Pair probabilistic language models with deterministic policy and data engines. That is the neurosymbolic pattern. The neuro part interprets. The symbolic part enforces. You can start today. If you wait, you risk becoming the next Nokia, whose leadership famously said they did not think they had done anything wrong, shortly before the business was gone.

End of transcript. The on-page text matches the .vtt and .srt files linked above.