25 AI engineers you should be following in 2026
Your X/LinkedIn feed is a product.
What you put in is what you get out, and if most of what you’re ingesting is founder bulletins and AI hype roundups, you’re basically working your skilled growth on vibes.
The engineers beneath are the individuals who really form how this expertise works: they publish the analysis, ship the tooling, break down the internals, and put up about it in ways in which reward cautious studying moderately than a fast dopamine hit.
The researchers whose posts learn like early-access papers
These are the individuals whose threads on X or posts on Substack commonly find yourself cited in different individuals’s papers, convention talks, and inside firm wikis.
25. Andrej Karpathy (@karpathy): Now at Anthropic working underneath Nick Joseph on pre-training, Karpathy joined in May 2026 after a stint constructing Eureka Labs. His nanochat undertaking has pushed the fee to duplicate GPT-2’s CORE benchmark rating all the way down to roughly $73 on a single 8xH100 node, a 600x discount over seven years.
24. Lilian Weng (@lilianweng): Her long-form posts on brokers, reasoning, and security are efficient reference documentation for the sphere. “LLM Powered Autonomous Agents” stays one of many most-cited posts in AI engineering, years after publication. Her writing on test-time compute is required studying if you work with reasoning fashions.
23. François Chollet (@fchollet): Creator of Keras, co-founder of Ndea (with Mike Knoop), and architect of the ARC Challenge, which carries a $1 million prize for real progress on summary reasoning. Chollet is unusually prepared to level out when scaling is producing diminishing returns moderately than the subsequent breakthrough. In a discipline full of individuals telling you the ceiling is infinite, that is genuinely helpful for calibration.
22. Nathan Lambert (@natolambert): His Interconnects e-newsletter is likely one of the clearest sources on post-training, RLHF, and reasoning fashions. When a brand new paper drops on reinforcement studying from verifiable rewards, Lambert’s breakdown is often the one engineers share.
21. Noam Brown (@polynoamial): Led the event of o1 and OpenAI’s reasoning mannequin line, by to the gold medal on the International Mathematical Olympiad in 2025. His threads on self-play and reinforcement studying utilized to LLMs are exact, dense, and value studying slowly.
20. Oriol Vinyals (@oriolvinyalsml): VP of Research and Deep Learning Lead at Google DeepThoughts, Gemini co-lead, and the particular person behind AlphaStar, AlphaFold, and WaveNet. If you need to perceive the place frontier multimodal architectures are heading, that is the feed.
19. Soumith Chintala (@soumithchintala): The unique drive behind PyTorch, now CTO of Thinking Machines Lab after leaving Meta in November 2025. Still codes. Still engages straight with open-source contributors. Still the type of one that makes you really feel barely unhealthy about what number of tabs you have open versus how a lot you’ve really shipped.
18. Shreya Shankar (@sh_reya): Research and analysis specialist whose work on knowledge pipelines and LLM judges is the closest factor the sphere has to a rigorous framework for understanding whether or not your system really works. Her paper “Who Validates the Validators?” is required studying earlier than you ship an LLM-as-judge setup.
The builders whose posts include precise code
This group posts much less about concept and extra about what they constructed, what failed, and what the system regarded like once they had been executed.
17. Simon Willison (@simonw): A developer who has been cataloging LLM habits since earlier than most firms had an AI technique. His 2026 PyCon lightning speak on six months of LLM developments was extensively shared as a good synthesis of what really modified. His immediate injection work is the canonical reference on instruction/knowledge separation, and his documentation habits are frankly embarrassing for the remainder of us, in the very best manner.
16. Swyx / Shawn Wang (@swyx): Co-host of the Latent Space podcast and e-newsletter, founding father of the AI Engineer World’s Fair (6,000+ engineers in 2026). He coined the phrase “AI engineer” as a definite position, and the sphere largely accepted it. His essays on what AI engineering really includes are extra helpful than most job descriptions.
15. Hamel Husain (@hamelhusain): His essay “Your AI Product Needs Evals” is the canonical start line for analysis infrastructure, and his follow-up “A Field Guide to Rapidly Improving AI Products” covers the total loop from evals to error evaluation to knowledge flywheels. Practical, opinionated, and nearly totally right.
14. Jason Liu (@jxnlco): Developer Experience Engineer on the Codex crew at OpenAI. Posts on RAG structure as a techniques drawback, instrument routing, and what structured outputs really allow past easy JSON extraction. His Instructor library has been adopted by 1000’s of manufacturing groups.
13. Abhishek Thakur (@abhi1thakur): Four-time Kaggle Grandmaster and the one who constructed AutoTrain at Hugging Face. His YouTube tutorials mix implementation rigor with real accessibility. If you want to grasp fine-tuning in apply, begin right here.
12. Sebastian Raschka (@rasbt): His e-newsletter covers LLM implementation and analysis with a degree of technical precision that rewards re-reading. Combines explanations with working PyTorch code constantly, which makes it genuinely completely different from most ML writing.
11. Lee Robinson (@leerob): VP of Developer Experience at Cursor, previously at Vercel. His tutorials on integrating coding agents into actual growth workflows are among the many most cited in the AI engineering neighborhood. Relevant in case your crew is deciding tips on how to work with Claude Code, Codex, or Cursor at scale.
10. Mira Murati (@miramurati): Former OpenAI CTO and founding father of Thinking Machines Lab (2025). One of the few individuals who has led groups delivery GPT-4, ChatGPT, and DALL-E, then began once more from scratch. Her perspective on safety-aligned industrial roadmaps is grounded in operational actuality.
9. Yohei Nakajima (@yoheinakajima): Creator of BabyAGI, one of many first public demonstrations of task-driven autonomous brokers. Now a VC at Untapped Capital. His posts on agent loop design stay virtually helpful even because the tooling round them has advanced considerably.
8. Eugene Yan (@eugeneyan): Applied science at Amazon and a meticulous blogger on manufacturing ML. His survey of task-specific LLM analysis strategies is likely one of the greatest sensible inventories of what really works per use case.
7. Cassie Kozyrkov (@quaesita): Founded the sphere of Decision Intelligence at Google. Her writing cuts by mannequin efficiency discussions to the precise query: what determination are you making an attempt to enhance, and does this metric seize it? Required following when your crew is about to select the fallacious metric once more. And sooner or later, each crew does.
6. Timnit Gebru (@timnitgebru): Computer scientist and main voice on algorithmic equity, bias, and the structural circumstances that produce unreliable AI systems. Her work is much less about building higher fashions and extra about interrogating what “higher” means. Worth following for the questions as a lot because the solutions.
5. Kate Crawford (@katecrawford): Author of *Atlas of AI* and co-director of the AI Now Institute at NYU. Her analysis on compute geography, labor circumstances, and knowledge provenance is the infrastructure context that the majority engineering discussions omit.
4. Jeremy Howard (@jeremyphoward): Co-founder of quick.ai and reply.ai, with a sustained dedication to creating frontier strategies accessible to practitioners who lack the sources of a significant lab. His latest work on small-team frontier engineering is more and more related as succesful open-source fashions develop into extensively out there.
3. Ethan Mollick (@emollick): Associate Professor at Wharton whose experiments on how LLMs have an effect on productiveness, creativity, and determination high quality are probably the most rigorous publicly out there work in that house. If your group is making an attempt to measure AI’s precise affect on data work, his papers are the benchmark.
2. Omar Khattab (@lateinteraction): Assistant Professor at MIT and creator of DSPy, the framework for programming LLM pipelines moderately than prompting them. His AI Engineer World’s Fair speak on constructing AI techniques that survive the bitter lesson is likely one of the higher architectural arguments for transferring previous brittle immediate engineering.
1. Dex Horthy (@dexhorthy): Founder of HumanLayer and creator of the 12-Factor Agents reference, which is likely one of the clearest design frameworks for manufacturing agent techniques. If your crew is delivery brokers and hitting the same old reliability issues, his posts will really feel like direct solutions.
A sensible observe on following
An inventory like this will flip right into a consumption drawback if you deal with it like a procuring cart. The signal-to-noise ratio on X degrades quick as soon as your feed fills up, so a greater strategy is to select a smaller cluster and browse the whole lot they put up for 30 days.
Figure out whose considering mannequin matches yours, and whose gaps complement yours.
Then increase. The objective is triangulation: when three individuals from completely different elements of this record reference the identical paper or shift in the identical week, that’s the sign value performing on.
One particular person saying one thing is attention-grabbing.
Three saying it independently means to clear your afternoon…
Meet AI builders in particular person
If you need to take the feed offline, the Agentic AI Summit Berlin on September 15, 2026, at The Ritz Carlton is constructed for precisely the practitioners this record represents: engineers and technical leaders delivery agentic techniques, with no expo-hall filler.
300+ attendees, centered periods, and the type of hallway conversations that really transfer issues ahead.
- Hands-on workshops and hackathons: Build and iterate on actual agent techniques alongside engineers dealing with the identical manufacturing challenges you are.
- Practitioner-only periods: Speakers from Hugging Face, NVIDIA, Siemens Energy, Lovable, and GetYourGuide masking what’s working in manufacturing proper now, throughout agent structure, MLOps, and utilized AI at scale.
- A community value preserving: The summit attracts senior technical leaders from throughout European AI, the type of room the place the conversations proceed effectively previous the closing keynote.
Early chook passes save €100.
