Today’s AI News

The episode covers multiple substantive topics with concrete technical details: Jev's cost structure ($42/billion tokens, 238x cheaper than Claude), the synthetic training mechanism for Kowa, and the five-level RSI framework. However, there is significant filler including extended metaphors (chef/oven analogy), celebratory asides ('That is wild'), and repetitive conversational scaffolding that dilutes the insight-to-minute ratio. The content is meaty but padded.

“JEV costs $42 for a billion input tokens...roughly equivalent to feeding the model a library of 10,000 average sized books for 42 bucks”

“They didn't touch a single piece of real customer data. The entire training set for Kowa is fully synthetic.”

The framing of AI shifting from 'chatty' to 'silent background infrastructure' is a useful lens, and the Jevons Paradox application is reasonably fresh. The synthetic training data approach for Kowa and the five-level RSI taxonomy from the Chinese paper provide novel structure. However, much of the underlying analysis (hallucination problems in LLMs, geopolitical AI divergence, privacy concerns with human feedback) circulates widely in AI discourse. The show repackages existing ideas effectively but offers limited counterintuitive or first-principles thinking.

“We are shifting from these chatty digital co workers to silent hyper fast background engines.”

“They didn't touch a single piece of real customer data...you set up a dynamic adversarial network using other frontier AI models”

The episode cites Diogo Almeida (ex-OpenAI researcher, Type Safe founder), implies knowledge of Salesforce's Kowa training, and references a Chinese academic paper, but neither speaker is directly quoted as a guest. The hosts are conversational journalists synthesizing third-party research and product announcements rather than interviewing practitioners. The content relies on secondary reporting rather than direct builder expertise. For a news/summary format this is acceptable, but it lacks the depth of direct operator perspective.

“Diogo Almeida. He's an ex OpenAI researcher who helped build the early methods that taught ChatGPT how to converse naturally with humans.”

“Over 30 Chinese AI researchers representing major institutions like ByteDance, Tsinghua University and Shanghai AI Lab just published a paper”

The episode excels at concrete numbers and named examples: Jev's $42/billion token pricing, 238x cost reduction vs Claude Fable, 70-500ms response times, three times fewer errors for Kowa on internal benchmarks, 75% of RSI papers at levels 1-2, less than 6% at level 5, Meta One pricing from $2.99 to $499/month, the Mopsy emergency care case. These are specific, quantified data points that ground the claims. Some sections (e.g., Gemini Extended Thinking, Odyssey 3 world models) lack hard metrics, but the overall specificity is high.

“JEV costs $42 for a billion input tokens...roughly 238 times cheaper to run, and it responds in 70 to 500 milliseconds”

“On their internal tests for tasks like updating deal stages or routing complex tickets, Kowa made three times fewer errors than the top Frontier models.”

The hosts demonstrate solid back-and-forth with genuine questions ('If it's not generating text, what is it actually doing?', 'Aren't they just scooping up all proprietary CRM data?') and logical follow-ups that build understanding progressively. However, the questioning is often gentle and affirming rather than challenging. There are few moments of productive disagreement or skepticism; the hosts largely guide listeners through a pre-built narrative. The Mopsy story feels like a predetermined heartwarming segment rather than organic discovery. The exchange is conversational but lacks the sharpness of a host pushing back or forcing guests/sources to defend claims rigorously.

“But if it's not generating text, what is it actually doing with those 10,000 books worth of input?”

“If Salesforce is training a new model to understand the deep nuances of sales and support, aren't they just scooping up all of their enterprise customers private CRM data to feed the machine?”