Inference-Time Scaling for Content Ideation:
Sampling Breadth, Thematic Coverage, and Novelty

An experiment report on generating short-form video ideas by independent sampling across five frontier language models, evaluated in embedding space against a labelled corpus of previously published work.

Note on this edition. Published on the AI Creator Adda site, with two changes from the original. Costs are omitted: the site prints no spend figures, so the comparison between models is stated in themes rather than dollars. And section 4.2 is corrected — an earlier draft reported a twelve-draw convergence across every vendor on the MIS Analyst role, which this run's own analysis.json does not support.

Abstract

A short-form creator publishing roughly one reel a day needs a continuous supply of ideas, and the supply is the bottleneck rather than the production. We tested whether sampling many independent draws from several frontier models, rather than iterating within one conversation, produces a usable and genuinely varied idea pool. One idea was drawn per API call, 30 calls per model across five models, each call entered through one of 20 rotating angles to prevent mode collapse. The 150 ideas were embedded with a 3072-dimensional model and clustered.

Every call returned a usable idea. The pool resolves into themes. Ideas sit further from the creator's existing work (mean cosine similarity to the nearest of 34 posted reels) than his own reels sit from each other (), so the generator is opening new ground rather than rediscovering the back catalogue. Separating the two effects the design confounds: at a fixed budget of 30 draws, spreading them across five models beats one model by themes ( against ), a real but modest gain, while raising the budget from 30 draws to 150 moves coverage from themes to . Scale mattered more than model mixture. We measured diversity and novelty, not quality; no human has yet rated whether these ideas perform.

Table 1. Summary of principal results. Each quantity is defined and derived in the section named.

QuantityValueWhere

1Why run this

The creator behind @your_data_scientist posts a teaching reel most days to an Indian audience learning AI and data science. A companion analysis of 35 reels posted between 28 June and 1 August 2026 found a median of 1,039 views against two breakouts at 34,405 and 58,264. The breakouts shared a shape: they reframed a job title the viewer already holds, and attached a one-word comment keyword to a real deliverable. The floor shared a different shape: eight reels between 458 and 672 views, five of which took exactly zero comments, each explaining a term or reporting a model launch without asking the viewer to decide anything.

That gives an unusually well-specified ideation problem. We know what the target looks like, we know the dominant failure mode, and we have a labelled corpus of prior attempts to measure novelty against. The open question is procedural: is it better to think hard once, or to sample widely and select?

This report contributes four things. A prompt built from measured performance rather than intuition, and a lens-rotation scheme that prevents the sampling collapse such a setup normally suffers. A measurement of how many distinct themes 150 draws actually yield. A novelty test against the creator's own back catalogue, with the baseline needed to make its number mean anything. And a separation of the two variables the design confounds, sampling scale and model mixture, which turns out to reverse the experiment's founding assumption.

2Method

2.1 The prompt

One prompt, roughly 5,700 characters, holding the creator's mission, the measured performance of all 35 prior reels with real view counts, and the rules distilled from them. Two choices in it are load-bearing. It leads with the losers' shared mechanic rather than the winners', because "five of the eight worst reels took zero comments" is a more correctable instruction than "career topics win". And it explicitly retires the winning sentence: the 58k reel's construction is named as used up, so the models reuse the mechanism underneath it instead of producing 150 variants of one headline.

2.2 Sampling

One idea per call, not a list of ten. A model asked for ten ideas optimises the set, padding it after the first strong entry; one idea per call makes each draw independent, which is what makes the 150 rows comparable and the saturation analysis meaningful.

Thirty identical prompts at temperature 1.0 do not explore a space. Models have strong modes and re-sample around them, so the draws collapse onto a handful of ideas in different words. Each call therefore enters through one of 20 rotating lenses, a required angle of attack (the failure mode, the verification problem, the career door outside software, the thing a senior will not tell a junior). Call n uses lens n mod 20, so every model covers the full set at least once.

2.3 Models and parameters

Five models through OpenRouter, 30 draws each, temperature 1.0, reasoning enabled, structured JSON output with a parse retry. Two configuration details cost real debugging and are worth recording. Setting provider.require_parameters, the pattern used elsewhere in this codebase to pin routing to schema-honouring providers, removes every viable endpoint for some vendors: Claude Opus 5 returned a 400 for unsupported structured outputs and GPT-5.6 Sol a 404 for no endpoints found. And reasoning tokens bill against max_tokens, so an initial ceiling of 1,200 was consumed entirely by DeepSeek's reasoning, returning empty content that looks exactly like a parse bug. The ceiling was raised to 6,000 after measuring Grok at 3,958 tokens on this prompt.

2.4 Measurement

All 150 ideas and all 34 recoverable prior reels were embedded with OpenAI text-embedding-3-large at 3072 dimensions, then row-normalised so a dot product is cosine similarity. Ideas were clustered with agglomerative clustering on cosine distance, average linkage, the cluster count chosen by silhouette over . Cluster labels come from one model call per cluster; exemplars are medoids, meaning a real idea nearest the cluster centre rather than a synthesised description. Every random component takes a fixed seed.

3Results

3.1 Yield

150 of 150 calls returned a parseable, non-empty idea.

3.2 Thematic structure

Silhouette selected clusters. The peak is shallow (, with neighbouring values within 0.004), which is normal for short-text embeddings and means the exact count should be read as "roughly this many themes", not as a discovered constant. Of these, contain ideas from more than one model and are singletons: one idea no other draw came near.

Figure 1.t-SNE projection (cosine metric, perplexity 25, seed fixed) of all 184 embedded texts. Each model has its own colour and its own marker shape, so a draw can be traced back to the model that made it; hollow rings are the reels already posted, and the numbered discs mark the centre of the 14 largest themes, keyed to Figure 5 and the table below it. Hover anything to read it. The five colours encode the models rather than the themes on purpose: 37 themes cannot be told apart by hue, whereas the model each draw came from is exactly the comparison worth making by eye. Read it for lean, not for territory. The colours overlap heavily, so no model owns a region, but each has a visible tilt: Claude Opus 5 sits denser toward the lower right, DeepSeek toward the left, and Gemini and Grok spread more evenly. That mild, real separation is the picture of the within-model versus across-model gap in 3.5 ( against ), and it is the mechanism behind the modest mixture gain in 4.1: a second vendor lands somewhere slightly different from where a 31st draw of the first would have. The two populations, generated and posted, are close to disjoint: the posted reels hold a compact region and the generated ideas largely sit outside it, which is the visual form of the novelty result quantified in 3.3. Two cautions. A 2D projection distorts distance and is shown for orientation only; every statistic in this paper is computed in the full 3072 dimensions, where the two flat principal components account for just of variance. And the two populations differ in text form as well as origin (a posted reel is embedded as its topic plus hook, an idea as a single sentence), so some of this visible separation may be style rather than substance. The baseline test in 3.3 is what rules that out, not this picture.

3.3 Novelty against the back catalogue

For each generated idea we took the cosine similarity to its nearest prior reel. The mean is , the maximum , the minimum .

That number is meaningless on its own. A similarity of could be far or near depending on how this embedding model scores any two texts from one narrow niche, so the reference that matters is how similar the creator's own reels are to each other. They are much more alike than any generated idea is to them.

Table 2. Nearest-neighbour cosine similarity within and between the two populations. The first row is the baseline that makes the third interpretable.

Nearest-neighbour cosine similarityMean

A posted reel's nearest neighbour among the other posted reels scores . A generated idea's nearest posted reel scores only . The generator is therefore working outside the creator's existing territory, by a margin larger than the territory's own internal spread. That is the result the raw number was trying to claim, and it survives the control.

The same table carries a caution the headline hides. Generated ideas are more similar to each other () than posted reels are to each other (). The pool is novel with respect to prior work while being internally tighter than the creator's real range. That is a predictable consequence of the prompt: 20 lenses all pointed at agentic systems, so the output is concentrated on-mission rather than spanning everything he actually posts about. Read as "a deep seam next to the existing catalogue", not "a broader catalogue".

Figure 2.Distribution of each generated idea's cosine similarity to its nearest of 34 posted reels. The whole mass sits between and , well below the ~0.8 region where texts read as restatements of each other.

3.4 Saturation, and whether 30 draws was the right number

This is the result the experiment was run to get. Resampling the draw order 400 times gives the expected number of distinct themes found after n draws.

Figure 3.Expected distinct themes after n draws, averaged over 400 random draw orders. The upper curve pools all five models; the lower is the mean of the five single-model curves, which is what 30 draws against one model actually buys. The gap at 30 draws is the entire case for using more than one model. The pooled curve is still rising at 150, but see the limitation in 5.3: sampling without replacement from a finite pool forces it to terminate at the true total, so the shape of the tail is partly an artifact and only the early slope should be read as a rate.

The marginal return falls sharply and legibly. The first 30 draws yield themes, about per draw. Draws 31 to 75 add more, about per draw. Draws 76 to 150 add , about per draw. A run half this size would have found roughly of the themes for half the money.

Read the two curves as answering different questions. The gap at 30 draws ( themes) is what mixing vendors buys at a fixed budget. The rise along the pooled curve is what buying more draws buys. The second is much larger than the first, which is the opposite of what this experiment was set up expecting; section 4.1 works through it.

3.5 Which model contributed what

Within-model similarity () exceeds across-model similarity (). The margin is small but consistent, and it means each model has a recognisable signature: its own draws resemble each other more than they resemble another vendor's. That is the mechanism behind the mixture gain in 4.1.

Note that the per-model totals below are each from 30 draws and are not comparable to the found by the pool, which had 150. They compare models to each other, not models to the pool.

Figure 4.Distinct themes each model reached in its own 30 draws. The spread between best () and worst () is five themes, so model choice moves coverage far less than budget does. The dashed line marks the themes the pool reached, but that came from 150 draws rather than 30 and is drawn for orientation, not as a fair comparison; see 4.1.

Table 3. Per-model contribution. Each model drew 30 ideas, so the theme counts are comparable to each other but not to the 37 found by the pooled 150 draws.

ModelThemes reachedInternal similaritySimilarity to prior work

3.5.1 Price and coverage are unrelated

DeepSeek V4 Flash reached themes. Claude Opus 5 reached . On this task, at this sample size, the least expensive model in the set returned more distinct themes than the most expensive one. If this experiment is repeated, the efficient configuration is to scale draws on a cheap model and keep one expensive model in the mix for contrast, rather than to spread the budget evenly.

3.6 The themes

The clusters, largest first. Sizes are uneven by a factor of ten: the head of the distribution is where several models converged on the same territory, and the tail is where a single draw found something on its own. Both ends are useful for different reasons, the head as consensus and the tail as reach.

Figure 5.Ideas per theme, largest first. Hover a bar for the exemplar and how many models reached it. The themes above the median hold of all ideas; the singletons at the bottom are ideas no other draw came near, which is where a wide sample earns its keep over a deep one.

Table 4. The 37 themes recovered by clustering, ordered by size. "Models" is the number of distinct models that reached the theme; the exemplar is the medoid, the real idea closest to the cluster centre.

#ThemeIdeasModelsExemplar (medoid)

4Discussion

4.1 Separating scale from mixture

The design confounds two variables, and the tempting summary ("one model reached themes, five models reached ") silently compares 30 draws against 150. Figure 3 separates them, and the honest accounting is less dramatic than that sentence.

Mixture, at fixed budget: 30 draws spread across five models find themes; 30 draws on one model find on average. Pooling buys themes, about more coverage for the same money. Real, worth having, not transformative.

Scale, at fixed mixture: going from 30 draws to 150 takes coverage from to themes. That is the larger effect by a wide margin.

So the headline is that scale mattered more than model choice, with mixture a secondary bonus. This inverts the intuition the experiment was designed around. If a budget is fixed and something must give, take more draws before taking more vendors. The practical caveat is that mixture is nearly free to add (it is a list of model ids) while scale is linear in cost, so there is no reason to give up the .

Why mixture helps at all is visible in the similarity numbers: within-model similarity () exceeds across-model similarity (). Each model has a signature and re-samples near it, so a second vendor lands in slightly different territory than a 31st draw from the first would.

4.2 Everything else

The lens rotation appears to be doing the work that a naive setup would have missed. A companion word-overlap analysis of the same 150 rows found zero cases of a model repeating itself, and all 20 lenses were represented in the output. Mode collapse was the expected failure of this design, and it did not occur. The cost of that insurance was writing 20 lines of prompt.

Where the models converged is informative rather than embarrassing. Two themes were reached independently by all five vendors: AI agent loop analogies (8 ideas) and AI Data Literacy (7). No model was pointed at either. Five vendors landing separately on the same ground suggests these are optima of the evidence itself rather than quirks of one model's training. Against that, 14 of the 37 themes came from a single model, which is the clearest argument in the run for not standardising on one vendor.

5Limitations

  1. We measured diversity and novelty, not quality. This is the important one. Nothing here shows these ideas will perform. Distinctness in embedding space is a necessary condition for a useful idea pool and nowhere near a sufficient one. The honest test is to shoot several and compare against the 1,039-view median, which has not been done.
  2. The cluster count is soft. A silhouette peak of is weak in absolute terms and the neighbouring candidates are within 0.004 of it. Read as an order of magnitude for the thematic range, not as a measured constant. Every downstream count that uses it inherits that softness.
  3. The saturation tail is partly an artifact. The curve resamples a finite pool without replacement, so it is forced to reach exactly the true total at n=150. Real additional draws would keep finding new themes at some low rate past that point. Only the early slope is a trustworthy rate; the flattening is real in direction but overstated in degree.
  4. One prompt, one topic, one day. All 150 draws share a single prompt on a single niche, run in one session. Nothing here establishes that the pooling advantage generalises to another domain, another prompt, or these same models next month.
  5. The pool is narrower than the creator's actual range. Generated ideas cluster more tightly around each other () than his posted reels do (). The 20 lenses all point at agentic systems, so this pool deepens one seam rather than widening the catalogue. A run intended to broaden topical range needs a different lens set, and this experiment says nothing about how that would perform.
  6. Embedding similarity is a proxy for sameness, not a definition of it. Two ideas with different wording and the same underlying premise can land far apart; two with shared vocabulary and different intent can land close. The clusters were spot-checked and read sensibly, but no human labelled all 150.
  7. Prior work is represented by 34 reels, not 35. One posted reel had no matched script and therefore no topic text to embed, so it is absent from every novelty comparison.

6Conclusion

The question was whether to think hard once or sample widely and select. On this task the answer is sample widely, with two corrections to the naive version of that advice.

The width that matters is draws, not vendors. The experiment was built on the assumption that mixing five frontier models was the lever. It is not. At a fixed budget of 30 draws, mixing buys extra themes over a single model ( against ), roughly . Multiplying the draws fivefold buys . Both are real; they are not the same size. Mixture survives as a recommendation only because it is nearly free to add, being a list of model ids, while draws cost money linearly.

The sampling needs structure or it does not scale. Thirty identical prompts at temperature 1 re-sample one region. The 20 rotating lenses cost twenty lines of prompt and produced zero within-model repetition across 150 draws, with every lens represented. Without that, this experiment would have measured the cost of rediscovering the same four ideas thirty times. Any attempt to replicate the yield reported here without a diversification scheme should expect a fraction of it.

What the pool is. distinct themes, sitting genuinely outside the creator's existing work: an idea is further from its nearest posted reel () than his reels are from each other (). It is a deep seam rather than a wider catalogue, because all 20 lenses point at agentic systems. For zero failed calls out of 150, that is a favourable trade against the alternative of a creator staring at a blank page.

What this does not establish. Every measurement here is of the pool's shape, not its worth. Distinctness and novelty are necessary conditions for a useful idea bank and nowhere near sufficient ones, and the selection half of "sample widely and select" was never tested: nothing in this work ranks the themes by expected performance. A generator that reliably produces novel, well-separated, useless ideas would score identically on every statistic in this report. Until reels get shot, the honest claim is that inference-time scaling solved the supply problem and left the selection problem open.

7What to do next

  1. Shoot four reels and measure them. The only test that matters. Take the top theme by convergence (MIS Analyst, reached independently by all five models), one mid-table theme, and two singletons from the tail, then compare against the 1,039-view median. Four data points will not be significant, but they separate "the pool is useful" from "the pool is novel and worthless", which no amount of further embedding analysis can.
  2. Re-run the budget split. Given that price bought no coverage here, the efficient configuration is most draws on the least expensive model plus one expensive model for contrast. Predicted by section 4.1: 150 draws weighted toward DeepSeek should reach a similar theme count. That is a cheap, falsifiable prediction and a direct test of whether the asymmetry holds.
  3. Widen the lenses to widen the pool. The current 20 all point at agentic systems, which is why the output is tighter than the creator's real range. A second lens set covering his other pillars would test whether the narrowness is a property of the method or just of this prompt.
  4. Add a selection stage. The experiment generated and stopped. A scoring pass against the measured performance rules, or a pairwise tournament, would close the loop from supply to shortlist, and could itself be evaluated against the creator's own ranking.
  5. Test whether the saturation point moves. The finite-pool artifact in 5.3 means the true rate past 150 draws is unknown. A 300-draw run would measure it directly and settle whether 150 was near the practical ceiling or nowhere near it.

8References and provenance

No formal literature review was carried out; this is an internal experiment report, not a survey, and it makes no claim to novelty against published work on sampling or ensembling. The methods below are standard and are cited so the choices can be checked rather than taken on trust.

  1. [1]L. van der Maaten and G. Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9:2579–2605, 2008. Basis for the projection in Figure 1, including the caution that it distorts distance.
  2. [2]P. J. Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20:53–65, 1987. Criterion used to select the cluster count, and the source of the caveat in Section 5.2 on shallow peaks.
  3. [3]F. Pedregosa et al. Scikit-learn: machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011. Version 1.9 supplied the agglomerative clustering (average linkage, cosine distance), the silhouette score, and the t-SNE implementation.
  4. [4]OpenAI. text-embedding-3-large, 3072 dimensions. Used for all embeddings in Section 3.
  5. [5]OpenRouter. Unified inference API. Used for the five generation models.

Model behaviour is provider-dependent and is not guaranteed stable over time, which bounds reproducibility of items [4] and [5] to the cached artifacts rather than to fresh calls.

One set of numbers in this document comes from outside the analysis pipeline and is stated as measured rather than derived: the prior-reel view counts, pulled from Instagram through the Apify Reel Scraper on 2 August 2026. View counts are a snapshot and keep climbing after collection.