♪ Take notes ♪The rules of honesty at this station.
One sentence of experience: the closer the protocol, the resolution, the status machine, the flow transfer, the format conversion code, the less likely it would be for the lower-level model to “get one running”. Once such codes are wrong, symptoms can occur at a very far upper level, and the cost of searching may be much higher than it would have been in advanced models at the outset.
A personal bug
One of my AI chat interface projects has a rendering problem. At first thought it was a rendering program that was premature and switched to a more mature Markdown/LaTeX rendering line, and the problem remained; AI once judged that there was a problem with big model output. It turns out that the real root cause is neither the rendering library nor the quality of the model, but the bottom SSE flow-to-pack logic is wrong.
Specifically, the unpacking function uses brackets to deal with the ” agent merges multiple SSE JSONs into one line ” , but does not judge whether the current character is inside the JSON string. So the LaTeX brackets in the body are also counted as structure brackets, spilling down depth, payload being sliced in the wrong position, JSON.parse being thrown wrong, the code being used to dump the whole delta silently. The closed brackets in the formula are the one that is dropped and the wrong formula is eventually rendered.
This bug is difficult to find in five layers of misleading: bottom error occurs when a package is broken; abnormally swallowed silently; mid-level expression is content deficit; upper layer is formulae rendering error; the outermost layer looks like “model output LaTeX instability”. The bottom protocol resolution is unreliable and all the models, rendering, filling and cleaning logic on the top are contaminated.
Core judgement
This is not a question of the merits of a model, but is a generic engineering judgement: low-level models can write codes that are low-risk, directly verifiable, and partially fail without contaminating the entire chain; but protocol resolution, flow-based incremental processing, various types of conversion, status machines, convergence, data migration, assurance payments, error processing and re-testing must be left to the design, realization and review of a strong model.
The risk of these codes is not that they “can’t be written”, but rather that they “look capable of running, but under border conditions quietly damage data”.
Risk-based allocation models: scripts, styles, batches for cheap models; generic business codes for medium-high models; protocols, resolution, status machines, data consistency, difficulties, bug rooting due to positioning to strong models. The key is not to “use expensive models only”, but to place the right-level model on the right level of risk.
The company uses the same AI: the less the better. Low-cost, low-risk tasks are pursued at a low cost and speed, with core links being generated by low-capacity models, subsequent miscalculation, back-to-work, and client losses likely to be more costly. AI Time managers need to learn to judge mission risks, not simply ask “what models are cheap”.