In an episode published November 20, 2024, Sara Hooker (Cohere) and Craig Wiley gave their GenAI forecasts for 2025 on Chain of Thought. I read that conversation as a dated spec rather than a scorecard: what two people deep in models, data, and production expected builders would need to do next. Several of their lines still read like work orders.
What changed after the one-click RAG era?
Wiley framed 2024 as the year RAG hype met production reality. Customers showed up with vendor demos that looked easy, then hit walls at scale. His line from that episode is the one I’d put on a wall:
prompt plus some docs doesn't equal a system that I can bet my company's reputation on.
He argued that systems that survive real stakes are often multi-component, compound, or agentic, and he asks customers about tuning, golden datasets, and LLM judges. None of that was a prediction about a single product release.
What to do with that today: treat any retrieval demo as phase zero. Before you promise a domain, write down what “wrong” costs (customer trust, compliance, revenue). If the cost is high, budget for compound architecture and eval infrastructure up front, not after the first executive demo.
Is “toolification” still the maturity test Wiley described?
In the same episode, Wiley tied industry direction to function calling: teams that use it aggressively look more mature to him. He pictured narrow, purpose-built capabilities orchestrated together instead of one company-wide megabot, and saw that tool-based approach as among the most promising paths to stepwise accuracy gains.
Wiley’s scope advice was blunt. A chatbot meant to cover “this massively broad area” makes evaluation schemas and domain understanding the bottleneck. Tractable slices (orders and delivery, then product specs, then troubleshooting) were his pattern for putting wins on the board.
Decision rule I’d use from that transcript: if you cannot name a golden dataset for one narrow intent, you are not ready to orchestrate five tools. Ship one tool-backed flow, measure it, then add neighbors.
Sketch (illustration only):
# Sketch: route to a narrow tool, not a general answerer
def handle_ticket(text: str, tools: dict[str, callable]) -> str:
intent = classify_intent(text) # your eval-backed classifier
if intent not in tools:
return escalate_to_human(text)
result = tools[intent](text)
if not passes_golden_checks(intent, result):
return escalate_to_human(text)
return result
How should Hooker’s dual trend (bigger and smaller) affect your model choices?
Hooker’s hot take on that episode was explicitly dual: models would get bigger, mainly for multimodal work where people push out the boat to see how far they can go, and a lot smaller through optimization in a broader way.
She tied the bigger path to multimodal work and uncertainty: humans use audio, visual, and motion to sense when something is off, and she expected algorithm work in that direction. She tied the smaller path to optimization across inference and stratified data.
For a builder, that split is a planning document. Multimodal and “push out the boat” work belongs where modality actually changes the task (PDFs with intertwined text and figures, claims packets, manuals with charts). Wiley’s wish on the same show was almost mundane by comparison: parse PDFs well enough that insurance claims docs stop being a science project. He said multimodal cost was still high enough that you should not default to it unless you need it.
Hooker also warned that tooling evaluation is messy because every API is shaped differently. You either get better at on-the-fly exploration or you saturate the data space for chosen tools. That is a data strategy question as much as a model pick.
When you need more examples for rare paths or privacy-blocked domains, synthetic data is the tool the glossary describes: generate to fill gaps, then validate against real data and solid evaluation. Do not treat volume as quality. Pair that with Hooker’s stratified-data angle: optimize which slices you train and serve, not only which checkpoint is biggest.
Where do preference alignment and eval fit if “trust” is the product?
Hooker spent part of the conversation on safety as multiple objectives: global norms plus local harms, and adapting models per jurisdiction without pretending one reward captures everyone. That connects to how RLHF actually works in practice: human rankings train a reward model, then the policy optimizes for it, with known limits (raters disagree, hard tasks are hard to judge, policies can game oversight).
She also stressed tracking progress after you tailor preferences, which sounds boring and is the kind of eval question Wiley said customers only started confronting in 2024. The November 2024 through-line from both guests is that production AI is eval-driven whether or not you feel ready.
Sketch for multi-objective monitoring (illustration only):
// Sketch: score outputs against explicit objectives, not one vague "quality"
type Objective = { name: string; weight: number; judge: (out: string) => number };
function releaseGate(output: string, objectives: Objective[], minTotal: number): boolean {
const total = objectives.reduce((sum, o) => sum + o.weight * o.judge(output), 0);
return total >= minTotal;
}
I'd map each production endpoint to a short objective list and refuse to widen scope until those scores are stable on a golden set.
Checklist
- Re-read Wiley’s November 2024 bar: reputation-grade systems are often compound and need tuning, golden data, and a plan for judging output; prompt plus docs is not enough.
- Use function calling and narrow tools as a scope discipline: one tractable intent, one eval schema, then orchestrate.
- Split Hooker’s bigger-vs-smaller bet: multimodal where modality is required; smaller optimized models and stratified (and carefully validated synthetic) data everywhere else.
- Treat preference and safety goals as measurable objectives, informed by how RLHF shapes models and where it fails.
- Default Wiley’s PDF pain into your roadmap if your users live in scanned docs, tables, and mixed layouts.
The full conversation, with the transcript, is on Chain of Thought.
Subscribe to the Chain of Thought newsletter for new episodes and write-ups like this one.
Drafted with AI assistance from the episode transcripts.
Top comments (0)