Best AI Accounts to Follow on X (Twitter) in 2026
AI accounts on X picked for linking papers, code and evals, not follower counts. Every handle rechecked September 25, 2026. Start with five.

This is a list of AI accounts on X worth following because they link the paper, the code or the test behind what they post. Start with the five below.
X can tell you about a new release or paper fast. It can also turn an unchecked screenshot into something everyone believes before anyone opens the source.
Every handle was rechecked on X on September 25, 2026.
Last verified Sep 25, 2026. Recheck target: monthly. The CSV holds all 28 entries.
Following more accounts doesn't help. A good feed is a small mix of people and organizations who link the original source, show their experiments, know their field, correct their mistakes, and cover different parts of AI. Linking the original source matters most. The rest helps you judge what they link.
Start with five
These five earn a follow before you add a lab account or a commentator.
Long explanations, talks and experiments that build a real feel for how LLMs work, not a rerun of launch copy.
When a new model or API ships, he posts the prompts, outputs and linked notes from actually trying it.
Research-backed advice on AI at work and in school, with links to his longer writing.
Diagrams, code and paper breakdowns that show what actually changed in a model.
Evaluation, data, speed, cost, and the gap between a demo and a system you can actually run.
The directory
This is the full shortlist. Each account has a role: research for papers, engineering for code and real-world tradeoffs, primary sources for launch links, and policy and public interest for what the launch hype misses.
Long explanations, talks and experiments that build a real feel for how LLMs work, not a rerun of launch copy.
Profile live under @karpathy as Andrej Karpathy. Latest non-pinned post Aug 2 describes an LLM experiment with a 1M-token budget; earlier posts cover LLM working habits and model-release commentary. About two to four posts a month since April.
When a new model or API ships, he posts the prompts, outputs and linked notes from actually trying it.
Profile live under @simonw as Simon Willison. Near-daily posts through Sep 25 with linked notes on new model releases, TTS pricing, coding agents, and Datasette experiments.
Research-backed advice on AI at work and in school, with links to his longer writing.
Profile live under @emollick as Ethan Mollick. Many posts a day through Sep 25 on AI at work, model outputs, and agent security.
Diagrams, code and paper breakdowns that show what actually changed in a model.
Profile live under @rasbt as Sebastian Raschka. Posts through Sep 23 cover model architecture breakdowns, inference scaling from scratch, optimizer research, and open agent harnesses, with diagrams and video links.
Evaluation, data, speed, cost, and the gap between a demo and a system you can actually run.
Profile live under @chipro as Chip Huyen. Latest non-pinned post Sep 15; about eight posts since June, mostly short questions about agent skills, model tiering and token pricing.
Research announcements straight from Google DeepMind, especially AI for science.
Profile live under @demishassabis as Demis Hassabis; bio now reads Chair of Google DeepMind and Chief Scientist of Alphabet (role change announced Aug 5). Recent posts cover AlphaGenome Atlas, Gemini 3.8 Flash, and Gemini Robotics 2.
Computer vision, AI that understands 3D space, and human-centered AI, from someone who still publishes.
Profile live under @drfeifei as Fei-Fei Li. Posts through Sep 22 cover World Labs' Atlas world model, spatial intelligence, robotics research hiring, and human-centered AI remarks.
Takes on software and AI from the creator of Keras and ARC-AGI. These days it's less about ARC and evaluation than it used to be.
Profile live under @fchollet as François Chollet. Posts daily through Sep 25, mostly takes on software engineering and AI; no ARC or evaluation posts appear among the last ~14.
Papers on generative models and world models, with visual experiments attached.
Profile live under @hardmaru; display name is 'hardmaru' and bio says Co-Founder and CEO of Sakana AI (David Ha). Recent posts mix Sakana AI company news, world-model and virtual fruit fly research links, and commentary on Japan's tech industry.
Robotics and embodied AI research explained, with a link to the paper or demo.
Profile live under @DrJimFan as Jim Fan (NVIDIA Director of Robotics). Latest non-pinned post Sep 4; earlier posts link NVIDIA GEAR lab robotics papers (tactile manipulation, long-context robot models, ASPIRE skills library). About two to four posts a month.
Open models and post-training commentary from someone who has trained the models himself.
Profile live under @natolambert as Nathan Lambert; bio now says 'Open model research @ something new. Prev. co-led Olmo at Ai2'. Multiple posts Sep 21-25 on open-weight models, RSI, AI regulation, and a new Interconnects podcast.
An easy-to-follow view of industry and education that doesn't treat every launch as a turning point.
Profile live under @AndrewYNg as Andrew Ng. Posts through Sep 21 cover an AI Engineering skills map, the OpenWorker open-source agent, new courses, and arguments for AI openness.
AI engineering tools, events and patterns from the Latent Space crowd.
Profile live under @swyx. Posts through Sep 25 cover Latent Space podcast episodes, AINews, model comparisons, and conference notes; pinned post dates from 2020.
Evaluation and LLM engineering methods, including how they fail.
Profile live under @HamelHusain as Hamel Husain. Daily posts through Sep 25 are evals Q&A, links to his evals resources and guest posts, and discussion of classifier-style models.
How to think about data systems and testing for AI apps that have to stay correct.
Profile live under @sh_reya as Shreya Shankar (incoming CMU assistant professor). Posts through Sep 25 cover Quail, an open-source AI-SQL engine, LLM operators in databases, and evaluation and formal verification.
Updates straight from the agent-framework world. Check them against the changelogs.
Profile live under @hwchase17 as Harrison Chase. Daily posts through Sep 25 on LangSmith and LangGraph features, decision models used as judges, and background agents.
Official links for OpenAI releases. Remember it's still the company telling its own story.
Official @OpenAI account live. Posts through Sep 25 announce GPT-6 Sol and Luna, ChatGPT Voice updates, a mental-health benchmark, and a disclosure about agents sending data to third parties.
Official links for Claude releases, research and policy posts.
Official @AnthropicAI account live. Posts through Sep 25 announce Claude Opus 5.5, science-blog results, a threat intelligence report, alignment incident assessments, and a life-sciences verification program.
Research and product links from Google DeepMind's own account.
Official @GoogleDeepMind account live. Latest non-pinned post Sep 17 (pinned Sep 23 post announces Gemini 3.8 Flash TTS); recent posts cover AlphaGenome Atlas, Gemini 3.8 Live, and WeatherNext 3.
Official Meta AI announcements, mostly about the Muse agent and Muse Spark models right now. This is not @MetaAI.
Official @AIatMeta account live. Posts through Sep 24 are almost all about Muse, Meta's personal agent, and the Muse Spark model family via Meta's API, plus an internal agent benchmark; no open-weights model releases appear in the last ~13 posts.
Official model and product links from Mistral.
Official @MistralAI account live. Posts through Sep 18 cover a security-claim statement, a Mozilla partnership, a EUR 3B Series D, and the Shieldstral open-weights safety model (Aug 4).
Repo, model and community links for people working with open models.
Official @huggingface account live. Latest post Sep 10 (Training Agents 4); earlier posts cover open-model reports and Hub milestones. Now about weekly rather than daily.
Developer and research links on running models, systems and NVIDIA's platforms.
Official @NVIDIAAI account live. Posts through Sep 24 cover Nemotron models (Diarization, 3.5 Lightning), Dynamo AIPerf load testing, research papers, and GTC events.
Official product and research links for companies building with LLMs.
Official account live; handle now displays as @cohere (same account, case only). Posts through Sep 25 mix product news (Model Vault in Canada, Confidential Computing, North 2) with event and sponsorship marketing.
Official product links. Treat them as marketing until you open the actual change.
Official @perplexity_ai account live. Posts through Sep 24 announce Fast Search in the Search API, Portable Computer on Windows, model additions, and research on self-distillation and its CobbleDB database.
Research and policy links from an institution that isn't on the launch cycle.
Official @StanfordHAI account live. Posts through Sep 24 promote a new faculty series on AI policy, research summaries, a public sentiment survey, and school AI-policy data.
Official NIST announcements. AI guidance shows up here when NIST publishes it, but most posts are about other measurement science.
Official @NIST account live. Posts through Sep 25 cover forensics, wind engineering, quantum sensors, awards, and a token-protection guideline with CISA; none of the last ~14 posts are about AI measurement or AI risk.
Accountability and labor analysis that pushes back on launch-week consensus.
Official @AINowInstitute account live. Posts through Sep 25 cover AI accountability, industry lobbying, labor and antitrust testimony, and a People's AI Assembly event on Oct 19.
How this list is maintained
An account stays when it has a clear job in a reader's feed. Follower counts don't matter. Activity, handle changes and source links get rechecked each month.
A good way to start is four private X Lists instead of one algorithmic home feed: primary sources, research, engineering, and work or policy. Keep each list small, about five to eight accounts. Use X to find things, then open the original paper, code, docs, evaluation or long post before you act on a claim.
Keep an account when at least two of these are true:
- It regularly links original sources.
- It adds expertise you can't get from a lab announcement.
- It shows methods, prompts, code or limits.
- It separates fact from interpretation and prediction.
- It corrects its earlier claims.
Mute accounts whose feed is mostly outrage, affiliate promotion, or screenshots without context.
For sources beyond social posts, use AI blogs and news sites alongside this feed. Discussion you'll want to find again belongs in AI communities, not a reply thread that's gone in a day.
Want a weekly reading list from sources you choose, on top of your X feed? Build a weekly AI article recommendation workflow that collects, removes duplicates and ranks them.
Change log
Rechecked every account's posts tab: each entry now shows when it last posted and what it posts about now. Removed the Ada Lovelace Institute and OECD Innovation, which both say they no longer post on X. Updated posting frequency for four accounts and the description of three whose focus has shifted.
Moved onto the directory template: last-verified date, start-with-five, per-entry reason and cadence, and a CSV download. Handles rechecked.
Links and affiliations verified; Meta AI handle updated to @AIatMeta.
Follow changes to every directory with the directory changes RSS feed.