TL;DR. If you only follow five: @karpathy for first-principles AI thinking [2][1], @simonw for hands-on LLM tooling and the daily blog [3][17], @rasbt (Sebastian Raschka) for from-scratch implementation [4][1], @_akhaliq (AK) for the arXiv firehose without the noise [6][33], and @GergelyOrosz for software-eng industry reality checks [18][20]. Skip the “AI influencer” lists ranked by follower count — they optimise for volume, not signal.
How to read this list
Curated for signal-to-noise, not reach. Most “top AI Twitter” articles rank on follower counts; that surfaces Sam Altman (4.2M) and Lex Fridman (4.2M) [16] — fine for headlines, weak for technical depth. The picks below are the ones repeatedly named by practitioners (KDnuggets [1], tkainrad’s “people who actually tweet about software engineering” [19], WeAreDevelopers [20], HyperWrite [29]) plus follow-graph evidence (e.g. Karpathy’s public endorsement of Willison [17]).
AI / LLM — research, practice, and curation
| Account | Who | Why follow | Beat |
|---|---|---|---|
| @karpathy [2] | Andrej Karpathy — ex-OpenAI/Tesla AI | Plain-English deep-learning intuition; threads on training, agents, “autoresearch” experiments. Top AI follow per multiple curators [1][29] | DL fundamentals, LLM intuition |
| @simonw [3] | Simon Willison — co-creator Django, builds llm CLI |
Daily working notes on prompts, tools, model releases; Karpathy publicly subscribes [17] | Practical LLM use, tooling |
| @rasbt [4] | Sebastian Raschka — author Build a Large Language Model (From Scratch) | Architecture breakdowns and implementation tutorials [1] | LLM internals, training |
| @_akhaliq [6] | AK — Hugging Face Papers curator | The most-referenced account for new arXiv drops; peer-reviewed paper found his shares measurably move citation counts [33] | arXiv triage |
| @dair_ai [5] | DAIR.AI org account | Paper-of-the-week threads, technical-but-readable explainers [1] | Research surveys |
| @askalphaxiv [11] | alphaXiv (project) | Social discussion layer over arXiv preprints [1] | Paper discussion |
| @TheAhmadOsman [7] | Ahmad Osman | AI infra, GPUs, running LLMs locally — niche but consistently technical [1] | Infra & local-LLM |
| @emollick [10] | Ethan Mollick — Wharton prof | Empirical, hands-on writeups of LLMs at work and in education [1] | LLMs in the wild |
| @natolambert [32] | Nathan Lambert — Ai2, Interconnects author | Open-model post-training, RLHF book; the deepest open-source-LLM beat | Open models, RLHF |
| @OfficialLoganK [21] | Logan Kilpatrick — Google DeepMind PM | Insider posts on Gemini API and developer tooling [29][30] | Google AI dev platform |
| @jeremyphoward [31] | Jeremy Howard — fast.ai co-founder | Practical-AI advocacy; opinionated education-focused threads [29] | Applied ML |
Big-name lab heads worth following but high-noise: @ylecun (Yann LeCun, Meta) [12], @demishassabis (DeepMind) [13], @AndrewYNg [14], @sama (OpenAI) [15]. Useful for company-strategy signals; less useful for technical content. Treat as headlines, not depth [16].
⚠ News-aggregator accounts like @TheRundownAI [9] and @mreflow (Matt Wolfe) [8] are listed by KDnuggets [1] and are fine if you want broad daily coverage — they trade depth for breadth. Don’t make them your primary feed.
Software engineering — practice, leadership, indie
| Account | Who | Why follow | Beat |
|---|---|---|---|
| @GergelyOrosz [18] | Gergely Orosz — The Pragmatic Engineer | #1 software-engineering Substack (1.1M+ subs); leaks, layoff-watch, hiring data, Big Tech eng-org analysis | Industry reality |
| @dhh [23] | David Heinemeier Hansson — Rails, 37signals | Strong opinions on framework design, cloud-exit, monoliths-vs-microservices; 436K followers [19] | Web frameworks, eng philosophy |
| @swyx [24] | Shawn Wang — Latent.Space, AI Engineer summit | Coined “AI Engineer”; bridges classical eng and AI tooling | AI × dev tooling |
| @levelsio [25] | Pieter Levels — Nomad List, Remote OK | Build-in-public indie-hacker; ship/revenue updates on real products [20] | Solo founder shipping |
| @steipete [22] | Peter Steinberger — ex-PSPDFKit, now OpenAI | Candid agentic-coding workflows; “the human side of building in the age of AI” [20] | Agentic coding |
| @addyosmani [26] | Addy Osmani — Google Chrome eng manager | Web perf, Core Web Vitals, frontend tooling [20] | Web performance |
| @dan_abramov [27] | Dan Abramov — ex-React core | Redux/CRA co-author; thoughtful long-form replies on UI architecture [19] | React internals |
| @kentcdodds [28] | Kent C. Dodds | Creator of React Testing Library; testing & full-stack patterns [20] | Testing, React |
Other names tkainrad’s “actually-tweets-about-engineering” list surfaces [19]: @bagder (Daniel Stenberg, curl), @mitsuhiko (Armin Ronacher, Flask/Sentry), @Nick_Craver (ex-Stack Overflow infra), @davidfowl (.NET architect), @codinghorror (Jeff Atwood, SO/Discourse), @troyhunt (HaveIBeenPwned). All low-volume, high-signal.
What to skip
- “Top 100 AI influencers” feedspot-style rankings [16] — sorted by reach, not signal. Useful only as a starting roster.
- Mega-accounts whose content has shifted to commentary/politics — even technically-credentialed accounts (LeCun, others) have drifted; mute, don’t unfollow, if you still want occasional company-strategy signal.
- AI-news aggregator accounts as your primary feed — they re-tweet what AK [6] and DAIR.AI [5] already surfaced first, with more noise.
Filter discipline
Two operational habits that matter more than which accounts you add:
- Use Lists, not the home feed. Build a private “AI signal” List with the top section above; pin its column.
- Mute aggressively. Mute words like “thread”, “🧵”, “BREAKING”, “this changes everything” — the noise accounts use them; the signal accounts don’t.
Caveats
- Follower counts above [16] are point-in-time (early 2026) and drift quickly.
- X profiles can go quiet — Karpathy went quiet for stretches in 2024 [2]; Simon Willison’s blog is more reliable than his X for the depth-first reader [17].
- For long-form depth, X is a discovery layer — pair this list with the bloggers/newsletter and YouTube shortlists from the same parent research.