Sub-second, reliably calibrated decision models turn AI from heavy conversational chatbots into a programmatic fourth logic gate—running inside DuckDB queries, UI keystrokes, and live web-grounded pipelines.
Conversational AI feels brilliant in chat because RLHF optimizes it to satisfy a human watching the screen. Drop that same model into unattended control flow, Diogo Almeida warns, and sporadic variance ruins automation: an unpredictable 5% failure rate forces a human reviewer back onto 100% of outputs.
“I often call it you get Einstein ninety five percent of the time, but Mr Bean five.”Diogo Almeida · I love you so jevving much · 00:08:57
Training specifically for narrow, dependable judgment turns model invocation from an unreliable assistant into a deterministic building block inside ordinary programs—like the blue LED transforming lighting rather than adding another filament:
“This is kind of conceptually adding a fourth semantic logic gate.”Diogo Almeida · I love you so jevving much · 00:12:12
Evidence: TypeSafe AI transcript 00:08:57, 00:12:12
Inspect most agent loops in production, Scott Roberts argues, and you find multi-second frontier LLMs burning tokens on simple branches that never needed open-ended generation:
“So one is there's too much intelligence.”Scott Roberts · Autocomplete for Actions · 00:00:37
“it's really just a conditional.”Scott Roberts · Autocomplete for Actions · 00:01:42
Exposing three zero-training primitives—probability booleans, up-to-255-way choices, and ordinal scoring—lets an application pick the right action card on every keystroke before a user ever hits Enter, while domain teams keep full control over the downstream schema:
“And you know we can do this at about twenty bucks for a million judgments.”Scott Roberts · Autocomplete for Actions · 00:06:42
Evidence: Workday transcript 00:00:37, 00:01:42, 00:06:42
Typical retrieval stacks encode documents into embeddings, store them in a vector database, run approximate similarity search, and feed candidates to a reranker that asks a yes/no question anyway. At moderate dataset sizes, Roberts skips the entire stack and asks the decision model directly:
“But with Jeff, we basically take out all the embedding steps.”Scott Roberts · Autocomplete for Actions · 00:08:03
Wired into DuckDB as a scalar SQL function, Jev screened 117 resumes in 2.3 seconds—catching semantic disqualifiers (like fraud-prevention work vs. actually committing fraud) that keyword and cosine retrieval miss:
“So this is against a DuckDB, where we have a function.”Scott Roberts · Autocomplete for Actions · 00:09:24
Evidence: Workday transcript 00:08:03, 00:09:24
Cloudflare became a model neolab for the same reason: its edge Workers needed fast System One routing and extraction models that run in a single forward pass with calibrated confidence scores, taking routine triage off heavy LLMs:
“We don't have to send every request to an LLM.”Michelle Chen · cloudflare is a neolab · 00:07:50
When the original closed-source System One model was unavailable, Cloudflare trained its own edge decision model (Clef) in a weekend and paired it with Clairic for open-ended extraction—proving the small-and-fast hypothesis only needed the right runtime:
“And we ended up training the model over 48 hours over the weekend, and then we launched it the following week.”Michelle Chen · cloudflare is a neolab · 00:09:53
Evidence: Cloudflare transcript 00:07:50, 00:09:53
Fast decision models still fail catastrophically when asked about current events outside their weights—returning stale guesses with misplaced certainty:
“And so it's kind of crazy that you would use a decision model in your workflow that has like almost no understanding of like what. And so this is where we think that like not only is it wrong, but also its confidence is too high.”Devin Gupta · The first Grounded Decision Model · 00:00:55
Parallel grounds each decision call against a ~100B-page live web index, turning guesswork into calibrated probabilities and source citations under one second at $1 per 1,000 requests—powering 100k-row market enrichment, KYC checks, and instant news routing:
“What is correct is much more confident that's correct if you use actual grounding to ground your decision. So there's like a whole set of things here that unlocks, and specifically like. Calibration is the important thing about good decision models.”Devin Gupta · The first Grounded Decision Model · 00:01:50
“It runs under a second in response time. It's one dollar per thousand requests. So. Two orders of magnitude, maybe an order of magnitude cheaper than the LLM workflow we were using to be back in April. And then it's also like an order of magnitude faster.”Devin Gupta · The first Grounded Decision Model · 00:03:01
“Not only do we give you the answer, we also give you the citation, so you can go fact check the source yourself.”Devin Gupta · The first Grounded Decision Model · 00:05:45
Evidence: Parallel transcript 00:00:55, 00:01:50, 00:03:01, 00:05:45
System One decision models are the conference’s cleanest uncontested architectural category. Every speaker converges on sub-second latency, typed non-conversational outputs, and calibrated confidence; talks diverge solely on where the call runs—inline in UI and DuckDB SQL, inside edge Workers, or backed by a live web index.
00:01:42, 00:07:50).00:08:03, 00:09:24).00:01:50, 00:05:45).if, switch, or filter step. Replacing slow chat completions with typed System One calls inside SQL, edge routing, or web-grounded APIs cuts latency and cost by 10–100× while eliminating fragile prompt parsing.A decision model fast and cheap enough to call everywhere replaces both LLM agent loops and retrieval stacks: it picks the right UI card on every keystroke and runs as a function inside a SQL query on every row.
Cloudflare became a model lab because small, fast System One models (Clef for decisions, Clairic for extraction) were the missing piece that fit its edge network and agent-cloud primitives, taking routine work off large LLMs.
Today's AI is too good to be true at assistance and too bad to be useful at automation because RLHF optimized it to please the human checking it; training for the right task, reliability, makes model calls a fourth logic gate inside ordinary software.
Grounding decision models with live web search from a large proprietary index solves uncalibrated overconfidence in ungrounded LLMs, delivering higher accuracy, verifiable citations, sub-second latency, and one-dollar-per-thousand-request pricing.