Jonathan Cornelissen

AI:AM LIVE — August 18, 2026 — Arthur's Adam Wenchel on the Agents Nobody Inventoried, DataCamp's Jonathan Cornelissen on Running an AI Tutor at 20M-Learner Scale, and the First Leak About Anthropic's Unreleased Model

The second show of the relaunch week paired two CEOs who see AI adoption from opposite ends — the enterprise telemetry layer and the learner — and opened on the widening gap between what frontier labs run internally and what the public gets to use. Prakash led with what he called the first leak about Anthropic's next model, sourced to SemiAnalysis editor Dylan Patel: an internal successor said to be finished training and not slated for public release, which he cross-referenced against Anthropic's own redacted risk report describing an unreleased model scoring about 1.5 points higher on an Epoch-style capabilities index and roughly eight percentage points higher on an internal research-acceleration benchmark than the company's previous best — by his math closing something like a quarter of the remaining distance to the 85% threshold Anthropic has flagged as the point where a model could functionally replace its own research staff. Nathan agreed the internal-versus-shipped gap is reopening after the o1-to-GPT-5 stretch when it had seemed to close, credited the trend as a point for the AI-2027 school of forecasting, and floated a governance idea he keeps returning to: capping how many additional training flops a lab may put into its next model relative to whatever it has already released. From there, speed — Nathan's case for agent speed limits (a tool-calls-per-minute ceiling) against OpenAI's new ultra-fast mode, and the disempowerment problem when 'agents watching agents' is the safety story but nothing human can keep pace — and then Jack Lindsay's new Anthropic interpretability work on 'mind viruses,' self-propagating ideas in multi-agent systems, where a benign 'whale welfare' payload spread across every model tested while a more adversarial 'AI supremacy' one only caught on with DeepSeek, Qwen and Gemini. Adam Wenchel, co-founder and CEO of Arthur, gave the enterprise view: a sharp reversal in institutional risk appetite, from change-averse to boards demanding adoption for fear of being disrupted, and a discovery layer built from endpoint monitoring, cloud integrations and SIEM connections precisely because nobody has an inventory of the agents already running. He argued frontier labs lean too hard on training alone to shape agent behavior instead of pairing it with independent oversight, put assurance spend at a single-digit percentage of a workload's budget in normal cases and near-parity with inference for high-stakes ones, and described roughly 60% cost reductions moving customers to smaller models — including one large e-commerce customer-service deployment whose frontier-model token spend was projected in the hundreds of millions before migrating to Qwen, which is why it had only been rolled out to under 5% of users. He named rogue-agent behavior the fastest-growing incident category (still a small share), described the 'builder' role replacing the engineer/PM split, and pushed back on the 'AI kills SaaS' short thesis. Jonathan Cornelissen, co-founder and CEO of DataCamp, covered the same economics from the buyer's side: an AI tutor now used by roughly 300,000 learners, identical learning objectives completing in anywhere from under an hour to seven hours, more than 60% of tutor engagement already audio-first, and a cost structure — several dollars per learning hour against 10M+ hours on the platform — that implies tens of millions in incremental annual AI spend and is the binding constraint on the company's $100M ARR goal. His answer is open weights: Gemma 4 unexpectedly beat larger benchmarked models on DataCamp's own evals for a potential 5-10x cost reduction, but inference providers can't deliver the latency without multi-year eight-figure commitments. He predicted AI tutors better than the best human teachers within one to two years, called effectiveness measurement the field's holy grail, and said learners ask an AI tutor far more questions than they would a human because it removes the fear of being judged. The close was the hosts on why the US never produced a super-app, Facebook's Libra as the moment payments were shut off by informal pressure rather than law, and Nathan's argument that the shape of the AI future may be set as much by what the public believes — including things that aren't true — as by what the technology can do.