Intelligence Snacks 73 - Model Explosion

Pete and Andy unpack a week where AI models seemed to multiply overnight: Anthropic updates, OpenAI’s cheaper models, Meta’s Muse, Grokbot, and the more interesting arrival of Jev. The big question is whether we need ever-bigger frontier models, or whether the real unlock is cheaper, faster, more specialised intelligence that can be sprinkled through workflows. The centre of the episode is Jev: a “system one” decision model built for fast classification, confidence scoring, routing and prioritisation. Pete and Andy explore why that matters for business processes, GraphRAG, tool calling, coding agents, fraud checks, refunds, traffic systems, learning graphs and app economics. Then they zoom out to consumer agents: Meta, WhatsApp, network effects, ad businesses, personal automation and whether people will hand their lives to Zuckerberg or choose to own their own agents. **Key Moments:** - [00:06] Welcome to episode 73, now with faces on YouTube. - [01:20] A big week in model land: Anthropic, OpenAI, pricing cuts and model-name weirdness. - [02:58] Are new models genuinely better, or are we comparing them against nerfed old ones? - [05:29] Pete argues for consistency, speed and cost over “more intelligence.” - [06:26] Jev enters the chat as the week’s most interesting model release. - [06:52] Andy tests Jev through OpenRouter on existing workflows. - [08:16] SEO signals, prioritisation and replacing vibe-based confidence scores. - [10:06] RLHF versus reinforcement learning via calibrated decisions. - [12:27] Why structured JSON decisions matter for automation. - [15:46] GraphRAG, context selection and cutting minutes of agent prep down to seconds. - [16:40] Refunds, fraud checks and business workflows where confidence changes the routing. - [18:28] Human escalation, graph traces and feedback loops for better decisions. - [19:33] The economics: why frontier models are overkill for simple sorting and routing. - [21:57] Jev as automation-native, not chat-native. - [23:45] Browser use is still painful when every mouse move costs time and tokens. - [24:08] Why Jev makes old ideas like Beacon feel interesting again. - [25:44] Open-source Jev-style models and the Intelligence Snacks thesis coming back around. - [27:35] IRL systems: traffic lights, waste, energy and transit routing. - [32:22] What happens to small language models when decision models exist? - [35:12] Intelligence Snack primitives: classification, prioritisation, translation and research. - [38:50] Jevons paradox and near-free judgment everywhere. - [41:47] “Fast is good”: why latency may matter more than bigger benchmarks. - [43:05] Learning graphs for codebases, agents and human understanding. - [47:29] Tool calling, agent loops and where expensive frontier reasoning is wasted. - [50:13] Code graphs, agent context and asking better internal questions. - [52:52] Meta Muse, Grokbot and the mainstreaming of personal agents. - [56:13] Agents versus consumer inertia: insurance, subscriptions and renegotiation. - [59:46] Do network effects beat capability? - [01:02:30] The Meta reservation: agents inside an ad business. - [01:04:43] Torrenting, distribution failures and extractive business models. - [01:08:27] Why own your agent instead of giving everything to Meta or Elon? - [01:12:21] Email is not really about communication; it’s about audit trails and accountability. **Friends of the Pod:** Jared Grigg, John, Justin Moon, Paul **Projects Mentioned:** Jev, TypeSafe, OpenRouter, Anthropic, OpenAI, Astra, Fable, Meta Muse, Grokbot, Instinct, Hugging Face, Beacon, GraphRAG, Autopilot, WhatsApp, Instagram, Amazon, Netflix **Concepts:** Decision models, system one models, calibrated confidence, RLHF, RLCD, model routing, tool calling, GraphRAG, small language models, Jevons paradox, local agents, consumer inertia, ad-driven AI **Quote:** “What you actually want out of these things is very, very quick inference.” — Pete