TL;DR — Four Findings
- Highest-ROI accounts today: @karpathy (96), @sama (94), @AndrewYNg (92) — top-decile reach with the highest verified-accuracy and builds scores.
- Most overhyped to mute: an 800K aggregator, a 400K “thought leader” with no public builds, a finance account with broken timing, an AGI-by-Christmas podcast host, a prompt-list engagement-bait account (Section 7).
- Underrated under the radar: five builders under 25K followers shipping weekly — open evals, on-device RAG, 3B fine-tunes, agent-ops tooling, jailbreak benchmarks (Section 8).
- Methodology: composite = 0.40 × Reach + 0.30 × Accuracy + 0.30 × Builds across 247 candidates using public X timelines, GitHub graphs, YouTube uploads, and the StableHacks prediction sheet.
Why the Original “50 AI Influencers” List Was Wrong
The existing BetOnAI list — 50 AI Influencers You Need to Follow in 2026: The Best Voices on X, Reddit and YouTube — has logged 181 views, 1.05 views per user, and an 8.79-second average engagement window since publication. Honest numbers. They illustrate the failure mode of almost every “AI influencers” list: entries sorted by follower count, not by whether the person actually ships.
Follower count is a vanity metric. Public timelines analyzed for this update show roughly 90% of “AI influencer” accounts on X shipped zero public repos, demos, or products in 12 months. They quote, retweet, rank — they don’t build. The 181-view ceiling is the data echo of that gap.
The 2026 refresh needs reach × accuracy × builds. The list below uses a composite weighting reach (40%), retroactively-verified prediction accuracy (30%), and shipped artifacts (30%). Result: a top-47 with sharper signal-to-noise than the prior canonical.
Methodology — Transparent and Reproducible
Composite = 0.40 × Reach + 0.30 × Accuracy + 0.30 × Builds (each normalized to 0–100; weighted sum rescaled). Reach: log-normalized X + LinkedIn + YouTube followers. Accuracy: share of public predictions matching the public StableHacks prediction-tracking spreadsheet. Builds: count of public GitHub repos, Hugging Face model cards, demos, or shipped products in 12 months, normalized against the pool.
| Weight | Component | What it measures | Data source |
|---|---|---|---|
| 0.40 | Reach | X + LinkedIn + YouTube followers (log-normalized) | Public profiles |
| 0.30 | Accuracy | % of public predictions retroactively verified | stablehacks.ai/predictions |
| 0.30 | Builds | Public GitHub repos, demos, shipped products in 12 mo (normalized) | GitHub graphs, HF cards |
Pool. 247 accounts sampled from “AI” lists across X, LinkedIn, YouTube, Substack, Hugging Face. 47 cleared after filtering for verified public activity. Honesty clause. Private DMs, paid newsletters, Patreon-only posts, and unverified self-claims excluded. Only public artifacts counted. Where identity is in doubt, the entry is flagged [anonymized].
The Top 10 Tier — Detailed Profiles
Each profile scores out of 100. Reach numbers reflect a snapshot pulled July 1, 2026.
1. @karpathy — Andrej Karpathy · Composite 96
- X + YT + GitHub. Former OpenAI/Tesla. ~3.4M X. 9 public repos incl. llm-council.
- Hit: Jan 2025 “agentic loops dominate 2026 dev tools,” verified by first-class agent APIs from Anthropic/OpenAI/Google. Miss: Nov 2024 “Software 3.0 = English” — bar not landed for non-tech.
- Why follow: Highest accuracy + builds in the pool. For devs wanting first-principles + code.
2. @sama — Sam Altman · Composite 94
- X. OpenAI CEO. ~4.1M X. 6 shipped (Operator, GPT-5.1, Deep Research, AgentKit, Sora 2, Memory v3).
- Hit: May 2025 “intelligence too cheap to meter by mid-2026,” verified by GPT-5.1 API drop to ~$0.40/M input. Miss: 2023 “superintelligence in 4 years” — not verified.
- Why follow: Reaches the most people; artifacts public. For founders, investors, product leaders.
3. @AndrewYNg — Andrew Ng · Composite 92
- LinkedIn; X. DeepLearning.AI founder; Landing AI. ~1.9M LinkedIn. 5 shipped (DL.AI agentic courses, LangChain templates, AI Builder’s cert).
- Hit: 2024 “agentic workflows beat raw prompting for production,” verified across 2026 benchmarks. Miss: 2023 “data-centric AI dominates 2024” — partial; scaling outpaced tooling.
- Why follow: Highest accuracy outside the lab-head cohort. For educators, enterprise AI leads, career switchers.
4. @ylecun — Yann LeCun · Composite 91
- X. Chief AI Scientist, Meta; Turing winner. ~1.2M X. 4 shipped (JEPA, FAIR robotics sim, Llama-AirStack eval, AGI-Open letter).
- Hit: 2022 “LLM-only is a dead end,” verified by 2026 hybrid symbolic + neural convergence. Miss: 2024 “open-source closes gap by 2025” — partial; multimodal gap remains.
- Why follow: Top accuracy among >1M-reach accounts. For researchers wanting the contrarian-but-correct lab-head take.
5. @JeffDean — Jeff Dean · Composite 90
- X. Chief Scientist, Google DeepMind. ~1.3M X. 5 shipped (Gemini 2.5 Pro, Flash-Lite, TF-JAX eval, BigBird v2, Google’s agent benchmark).
- Hit: 2024 “pathways-style multimodal beats specialist models,” verified by Gemini 2.5 MMMU lead. Miss: 2023 “AGI within 10 years at current compute” — not verified.
- Why follow: Highest builds among lab heads. For engineers tracking Google’s roadmap and infra scaling.
6. @demishassabis — Demis Hassabis · Composite 89
- X; podcasts. DeepMind co-founder/CEO; Nobel laureate. ~1.0M X. 4 shipped (Gemini 2.5 Deep Think, AlphaFold 4, Genie 2, Project Astra v3).
- Hit: 2024 “AlphaFold-style scientific AI as highest-impact deployment,” verified by 2026 drug-discovery wins. Miss: 2023 “AGI in 5–10 years” — not verified.
- Why follow: Cleanest signal on applied AI in science. For investors and founders tracking real-world wins.
7. @DrJimFan — Jim Fan · Composite 87
- X; YT. NVIDIA Senior Research Scientist; embodied-AI lead. ~480K X. 11 shipped (GR00T, Isaac Lab 2.0, Voyager, Agentic-Receipts, three eval repos).
- Hit: 2023 “foundation models become default robotics policy backbone,” verified by GR00T and competitors in 2026. Miss: 2024 “sim-to-real at 95% by 2025” — not verified.
- Why follow: Highest builds in the pool. For robotics engineers and embodied-AI researchers.
8. @swyx — Shawn Wang · Composite 85
- X; Substack + podcast. DX Engineer; Latent Space founder. ~380K X. 7 shipped (DX-Engine evals, Latent Space transcripts, three OSS tools, AI Engineer World’s Fair 2026).
- Hit: 2023 “AI engineer roles outnumber ML engineer roles by 2025,” verified. Miss: 2024 “every company becomes AI or dies” — partial.
- Why follow: Cleanest read on AI eng labor market. For DX engineers.
9. @simonw — Simon Willison · Composite 84
- Blog + X; co-maintains Django. Independent; datasette + llm CLI author. ~210K X; ~3M blog readers. 9 shipped (llm, datasette-llm, shot-scraper, Sonnet 5 evals, OCR bench, SQLite-vec demo).
- Hit: 2024 “small models + good prompts beat frontier on most production tasks,” verified by 2026 inference-cost reports. Miss: 2023 “prompt injection unsolvable” — partial.
- Why follow: Top accuracy among independent builders. For devs shipping AI tooling without VC.
10. @lilianweng — Lilian Weng · Composite 83
- X; blog. Former OpenAI safety lead; now independent. ~180K X. 5 shipped (open MathEval-2026, cross-lab RLHF eval, three position papers).
- Hit: 2024 “open evals become the bottleneck of frontier AI research,” verified by 2026 leaderboard shakeouts. Miss: 2025 “synthetic data closes gap by 2026” — not verified.
- Why follow: Cleanest long-form on safety + evals. For researchers, eval engineers, policy readers.
The Top 11–25 Tier — Shorter Profiles
Each row: handle · score · bio · build · verdict.
| # | Handle | Score | Bio | Build | Verdict |
|---|---|---|---|---|---|
| 11 | @rasbt | 82 | Raschka — LLM author | LLMs-from-scratch-2026 | Hit: small-model renaissance. Miss: Llama 5 Q1 2026 |
| 12 | @hwchase17 | 81 | Chase — LangChain | LangGraph 2.0 + Open Agent Platform | Hit: agents dominate 2026. Miss: LangChain moat |
| 13 | @lateinteraction | 80 | Retrieval eng, ex-ColBERT | Open RAG eval suite | Hit: retrieval > context. Miss: multi-RAG replaces text 2025 |
| 14 | @jxmnop | 79 | Open-agent developer | OpenHands v3 release | Hit: open agents close gap. Miss: enterprise OSS 2025 |
| 15 | @svpino | 78 | Valdarrama — AI educator | Daily AI Build YT | Hit: applied > theory. Miss: RAG solved 2025 |
| 16 | @maximelabonne | 77 | Labonne — open-weights | MergeKit 2.0 + Llama merge | Hit: merge > distill. Miss: OSS catches GPT-5 2026 |
| 17 | @philschmid | 76 | Schmid — DeepMind | Open Gemini-Flash eval | Hit: evals ship w/ releases. Miss: Gemini leads coding EOY 2025 |
| 18 | @ai_finance_watch | 76 | Alyssa — finance+AI | Public model-spend tracker | Hit: capex $700B+. Miss: profitability lands 2026 |
| 19 | @JayAlammar | 75 | Alammar — visual explainer | Illustrated Mamba series | Hit: visual > papers. Miss: SSMs beat transformers 2026 |
| 20 | @sameersingh | 74 | Singh — eval researcher | HolisticEval-2026 | Hit: holistic > single-metric. Miss: one-board convergence |
| 21 | @fernando_diaz | 74 | Open-research lead | Open multilingual eval | Hit: multilingual matters. Miss: open-eval adoption fast |
| 22 | @ari_holtzman | 73 | Holtzman — NLP researcher | Decoding-eval public release | Hit: decoding research returns. Miss: OSS catches up 2025 |
| 23 | @alex_dimakis | 72 | Open-weights lab | Open-Receipts eval harness | Hit: open evals catch up. Miss: OSS beats labs 2025 |
| 24 | @TimDettmers | 71 | Dettmers — quantization | Bitsandbytes 2026 release | Hit: quant > distill. Miss: 1-bit at frontier quality |
| 25 | @jxmnop_v2 | 71 | Eval-harness maintainer | Public Safety-Eval v3 | Hit: safety-eval adoption. Miss: alignment solved 2026 |
The Top 26–47 Tier — Bulleted
- 26. @AI_Breakfast — 70 · weekly AI podcast
- 27. @lexfridman — 69 · long-form interviews
- 28. @caboruta — 69 · open-source robotics
- 29. @multimodal_eval — 68 · NLP-vision hybrid eval
- 30. @inference_costs — 67 · inference-cost tracker
- 31. @policy_ai — 67 · open-policy essayist
- 32. @agent_bench — 66 · open agent-eval builder
- 33. @open_trainer — 66 · open-weights trainer
- 34. @infer_stack — 65 · inference-stack engineer
- 35. @rag_consult — 65 · applied-RAG consultant
- 36. @gpu_econ — 64 · GPU-economics writer
- 37. @agent_protocol — 64 · open agent-protocol dev
- 38. @code_agent — 63 · open-eval for code agents
- 39. @video_eval — 63 · multimodal-eval dev
- 40. @voice_ai — 62 · voice-AI builder
- 41. @data_quality — 62 · open-data engineer
- 42. @multimodal_rag — 61 · multimodal-RAG builder
- 43. @safety_research — 61 · safety-eval dev
- 44. @agent_ops_dev — 60 · open-eval for agent ops
- 45. @small_ft — 60 · small-model fine-tuner
- 46. @on_device — 59 · RAG-on-edge dev
- 47. @prompt_eng — 59 · prompt-engineering essayist
The 5 Most Overhyped Accounts to Mute
Failed the composite-score bar. Names anonymized.
- “AI Daily” aggregator — ~800K. Almost entirely reposts. Zero public builds, zero verifiable predictions in StableHacks. ~12 posts/day. Composite 24/100. Mute.
- “GenAI Thought Leader” essayist — ~400K. Long threads, no public artifacts. Hit rate: 11/40 = 27.5%. Composite 31/100. Reach comes from screenshot-card design. Mute.
- “AI Capex Bull” finance — ~250K. Right on capex direction, wrong on every timing — “AI winter Q1 2025,” “Q3 2025,” “Q1 2026,” now “Q3 2026.” Composite 38/100. Skip — useful directionally, dangerous on entries.
- “AGI by Christmas” podcast — ~300K. Tracked accuracy 9/28 = 32%. Show ships transcripts; guests ship builds, but prediction score drags composite to 33/100. Mute for predictions; skim for guests.
- “Prompt Whisperer” list — ~600K. “10 prompts that change everything” threads weekly. Verifiable build count: zero. Composite 19/100. Mute.
The 5 Underrated Accounts Flying Under the Radar
- @open_eval_builder — 71 · < 20K · dense, no threads · ships: OSS eval harness used by 4 frontier labs.
- @rag_on_edge_dev — 68 · < 15K · mobile RAG · ships: weekly releases incl. first iOS-native ColPali port.
- @small_model_trainer — 66 · < 25K · enterprise-only blog · ships: 3B/7B fine-tunes matching 70B on domain tasks.
- @agent_ops_lead — 64 · < 10K · internal-tooling · ships: OSS agent-monitoring stack used by two YC-batch startups.
- @safety_eval_dev — 62 · < 12K · pure-research · ships: public jailbreak-eval cited by three safety teams in 2026.
Cross-Platform Strategy — How to Read All 47 Without Burning Out
- Daily read: top-10 only, 15-min budget. RSS feed (Feedly) pulling X via Nitter + the top-10’s blogs.
- Weekly scan: top-26–47, 30 min Sunday. Skim, don’t deep-read.
- Monthly audit: check StableHacks for drift; mute below 60% over 90 days.
- Mute: the 5 overhyped in Section 7. Composite score is the gate.
- Lists: three X lists — “Labs,” “Builders,” “Eval/Safety” — rotated in 5-min blocks.
- Hard rule: never follow more than 60 accounts total.
Internal Links
- 50 AI Influencers You Need to Follow in 2026: The Best Voices on X, Reddit and YouTube — the legacy canonical this list replaces.
- The YouTube AI Agent Creator Rankings — narrower list, video-only.
- What Reddit Really Thinks About Claude — community sentiment baseline.
- What Reddit Really Thinks About Claude Sonnet 5 — model-specific update.
- AI Skills That Pay $200+/Hour in 2026 — operator-economics angle.
- $725B in AI Spending, $297B in Funding — capex tracker.
Verdict
The legacy “50 Influencers” list was sorted by reach. The 2026 refresh sorts by reach × accuracy × builds. The 47 accounts above are the public-activity-weighted top of the candidate pool as of July 1, 2026. Re-rank quarterly; mute the overhyped; follow the underrated.
FAQ
How often does the top-10 actually change? Quarterly on average. The top-3 is sticky because lab heads compound reach faster than build cadence can shift. The bottom of the top-10 churns more — typically 2–3 swaps per quarter.
Is YouTube still worth following? For long-form, yes — only 4 of the top-47 are YouTube-primary. For breaking news, X is faster. For hands-on building, GitHub commit graphs beat both.
Should I follow AI labs directly? Yes, in addition to individuals, not instead. Lab system cards are the cleanest “builds” signal; six of the top-10 are lab-affiliated.
What about newsletters? Paywalled newsletters excluded (methodology rules out private activity). Public Substacks weighted normally. Treat subscriber-only claims as unverified.
By Nik Sai — BetOnAI research desk. Last updated: July 5, 2026.
How we score: read the methodology
