The 50 Most Active AI Founders on X,Composite-Ranked Q2 2026

Inputs and weights

FORKOFF combines 4 first-party X engagement inputs into the composite-rank under two parallel weight schemes published simultaneously on every founder card. The PRIMARY equal-weight scheme avoids the reviewer-collapse risk of operator-picked weights. The SECONDARY platform-aware scheme reweights toward engagement rate (the highest-signal X metric per X's 2023 open-sourced ranking algorithm) so readers can audit how the rank shifts.

composite_score_primary =
  0.25 * rank_normalized(posts_per_week)
+ 0.25 * rank_normalized(avg_engagement_rate)
+ 0.25 * rank_normalized(follower_delta_30d)
+ 0.25 * rank_normalized(reply_rate)

composite_score_secondary =
  0.25 * rank_normalized(posts_per_week)
+ 0.35 * rank_normalized(avg_engagement_rate)
+ 0.20 * rank_normalized(follower_delta_30d)
+ 0.20 * rank_normalized(reply_rate)

7 acknowledged biases of the sample

1. Survivor bias. 53 handles were hand-curated by the FORKOFF editorial team and are not a random sample of all AI founders on X. The ranking reflects the cohort, not the universe.
2. Recency bias. 30-day window exposure decays between day 1 and day 31 of the snapshot. Quarterly refresh resets the window.
3. Platform bias. X-only by design. LinkedIn-heavy founders appear weaker than they are in the broader AI conversation.
4. Language bias. English-language gate excludes influential non-English AI communities, especially Chinese-language and Japanese-language ecosystems.
5. Operator self-exclusion. No FORKOFF team member or client appears. Disclosed for transparency.
6. Self-selection bias. Only founders who chose to be active on X are observable. Silent founders are invisible regardless of company impact, so the cohort over-represents the operator-archetype that already commits to public posting and under-represents heads-down builders.
7. Engagement-rate ceiling at mega-tier. Accounts above roughly 5M followers face a structural engagement-rate ceiling unrelated to content quality (audience-passivity at scale). The ratio compresses toward zero as follower count climbs, which artificially penalizes the largest accounts under the rank-normalized engagement-rate input.

Resampling protocol (cluster-bootstrap, founder-as-cluster)

The bootstrap unit is the founder. For each founder, the 30-day post-level data is the cluster, and the resampling-with-replacement operates on individual posts inside that cluster. The verbatim algorithm runs as follows for every founder f in the cohort.

1. Pull the 30-day post-level dataset P_f for founder f (n posts).
2. Draw a resample P_f* of size n from P_f with replacement.
3. Recompute the 4 input metrics (posts_per_week, avg_engagement_rate, follower_delta_30d_normalized, reply_rate) on P_f*.
4. Hold the cohort as a fixed reference distribution. Re-rank f within the cohort using the resampled metrics and the locked PRIMARY weights.
5. Record the resampled rank R_f* for this iteration.
6. Repeat steps 2 to 5 for B=1000 iterations.
7. The 2.5 percentile and 97.5 percentile of the {R_f*} distribution form the published 95% CI bracket.

Cluster-bootstrap with founder-as-cluster (per Davison and Hinkley 1997) preserves the within-founder correlation structure that an iid post-level resample would destroy. Posts from the same founder share account-level systematic drivers (follower count, niche, time-of-day pattern); resampling at the founder level retains that structure while still capturing post-level noise inside the cluster.

Multiple-comparisons correction (BH-FDR)

Pairwise rank-separation tests on neighboring founders use the Benjamini-Hochberg false-discovery-rate correction at alpha=0.05. With m pairwise tests sorted by ascending p-value, the procedure rejects every test i where p_(i) is at most (i / m) * alpha. The correction holds the expected proportion of false positives among rejected nulls at or below 5%.

Uncorrected pairwise tests at alpha=0.05 would expect roughly 12 false positives across the 245 within-5-ranks pairs in a 53-founder cohort. BH-FDR keeps the false-discovery rate at the published 5% floor and surfaces only the rank-separations that survive the correction. Tied rank-separations (p > 0.05 after correction) are flagged on the expanded ranks-11-50 row footer for the reader.

Bayesian posterior fallback for sparse founders

Founders with fewer than 10 posts in the 30-day window swap the bootstrap point estimate for a Bayesian posterior mean on the engagement-rate and reply-rate inputs. The prior is Beta(1, 1), equivalent to Uniform(0, 1). The posterior is Beta(1 + successes, 1 + failures), and the posterior mean ((1 + successes) / (2 + n)) replaces the raw rate.

The Beta(1, 1) prior is the maximum-entropy choice over the (0, 1) support and adds minimal information to the data. The posterior mean shrinks the raw rate toward 0.5 by an amount proportional to the inverse of the sample size, which is the desired behavior for sparse counts where the raw point estimate is structurally unstable. Footer-flag on the founder card discloses Bayesian fallback per founder where it applies.

What the CI does NOT capture

The published 95% CI captures within-founder sampling noise on post-level metrics. It does not capture model uncertainty on weight choice or sampling uncertainty on cohort selection.

The CI assumes the PRIMARY 0.25/0.25/0.25/0.25 weights are fixed, not random; weight uncertainty would widen every bracket by a factor that depends on how far the operator priors sit from equal-weight. The CI also assumes the 53-founder cohort is the population of interest, not a sample from a larger universe; cohort-selection uncertainty would widen the brackets further on the rank-position scale. Readers who want the joint uncertainty over weight choice plus cohort drift should re-rank against the downloadable CSV under their own model and aggregate the resulting rank distribution.

Rank-based normalization

FORKOFF applies rank-normalization (percentile rank divided by n=53) on every input rather than min-max to preserve interpretability on heavy-tail follower distributions. Min-max compresses 99% of the cohort into a near-zero band when a single outlier (Elon at 200M followers) anchors the upper bound. Rank-based normalization is outlier-robust by construction and stable across bootstrap resamples.

Bootstrap procedure

FORKOFF runs 1000-iteration cluster-bootstrap with founder-as-cluster per Davison and Hinkley 1997 to produce 95% CI on each rank position. Each founder 30-day post-level data is resampled with replacement (n = posts in window). The 4 input metrics are recomputed on the resample. The cohort is held as a fixed reference distribution and the founder composite is re-normalized and re-ranked within it. The 2.5 percentile and 97.5 percentile of the per-founder rank distribution become the published 95% CI bracket.

Founders with fewer than 10 posts in the window use a Bayesian posterior with a weak Beta(1,1) prior on the engagement-rate and reply-rate inputs. Paired bootstrap on any two founders within 5 ranks produces a Benjamini-Hochberg FDR-corrected p-value on rank separation; ties at p > 0.05 are flagged in the dropdown row footer.

Sample frame and self-exclusion

FORKOFF locks the cohort at 53 hand-curated handles across 6 categories with 4 backup handles in seed-YAML order for failover. Inclusion: verified handle returning HTTP 200, founder-level role at a meaningful AI company, English-language 80% or more in window, 1000 followers or more, 4 original posts or more in window. Exclusion: bot heuristics, suspended, private, dormant, sub-1000 followers, non-English-majority, operator-flagged conflicts of interest.

No FORKOFF team member or named FORKOFF client appears in the cohort. The editorial self-exclusion firewall is the canonical answer to the listicle-self-rank attack vector that compromises most agency listicles.

Tie-breakers and floor

1. If composite scores tie to 3 decimals, higher avg_engagement_rate wins (quality over cadence).
2. If still tied, higher reply_rate wins (operator-grade beats broadcaster).
3. If still tied, lower followers_end wins (smaller account is more impressive at the same composite).
4. If still tied, alphabetical by handle with a footnote.
5. Bottom-quartile floor: founders below the 25th percentile composite drop from the published list with reason logged.

Open dataset and methodology PDF

Required attribution: FORKOFF, Top 50 AI Founders Most Active on X, Q2 2026 snapshot. Methodology cross-references the cold-email open-rate methodology and the ASVC discovery-gap methodology.