Andrej Karpathy

“Author of "Andrej Karpathy Discusses Using GPT-5.1 to Evaluate Past HN Predictions in His New Blog Post"”

“Author of "Andrej Karpathy Finds Python's Random.seed Removes the Sign Bit, Causing Negative and Positive Seeds"”

Large language models represent a fundamentally different form of intelligence than animal intelligence, shaped by commercial optimization rather than biological evolution, and understanding this distinction is critical for reasoning about AI systems.

“Author of "The space of intelligences is large" in Twitter/X”

AI represents a new computing paradigm (Software 2.0) where the most predictive feature of task automation is verifiability rather than algorithmic specificity, fundamentally reshaping which jobs and tasks are amenable to automation.

“Author of "AI Is a New Computing Paradigm" in Thread Reader App”

Moltbook, a social network for AI agents that went viral with 1.7 million bot accounts, was ultimately 'AI theater' rather than a genuine glimpse of autonomous agent futures, with much of its activity being spam, crypto scams, and human-generated performance.

“OpenAI cofounder and influential AI researcher who called Moltbook 'the most incredible sci-fi takeoff-adjacent thing' on X.”

AI startups should optimize inference spending as a growth investment and viral distribution mechanism rather than as a margin cost, treating compute as CAC replacement to drive product-led growth.

“Quoted describing Cursor's 'vibe coding' experience as a key innovation driver”

AI startups winning the market are agent labs that ship products solving real problems using existing frontier models, rather than model labs spending years and billions on foundation model R&D.

“Quoted on industry overshooting tooling relative to present capability and the importance of integration work”

Andrej Karpathy argues the 'decade of agents' is more accurate than 2025 being the 'year of agents,' citing insufficient groundwork in intelligence and context handling, with AI agent adoption likely peaking in impact by 2027-2028.

“Discusses AI agents, AGI timeline predictions, and historical shifts in AI research over 15 years of experience”

OpenAI's o1 model introduces test-time compute as a new paradigm for AI improvement by training models via reinforcement learning to perform chain of thought reasoning natively, enabling performance gains through longer inference time rather than just larger training runs.

“Joked on X that o1 refuses to solve the Riemann Hypothesis”

Large LLM context windows degrade model performance and increase costs significantly; effective context management ('context engineering') is essential to optimize both performance and API expenses.

“Ex-OpenAI researcher quoted defining context engineering as 'the delicate art and science of filling the context window with just the right information.'”

SkyPilot v0.11 provides a unified interface for AI teams to run, manage, and scale workloads across multiple cloud providers and on-premises infrastructure with advanced scheduling, cost optimization, and enterprise-grade features.

“Created nanochat, described as the best ChatGPT that $100 can buy, which can be trained and served using SkyPilot.”

Effective agent design patterns center on context management as a finite resource, with successful agents using computer access, hierarchical action spaces, and deliberately limited tool sets to optimize long-running autonomous capabilities.

“Framed the need for context engineering as 'the delicate art and science of filling the context window with just the right information.'”

Reinforcement Learning from Verifiable Rewards (RLVR) emerged as a major new training paradigm in 2025, enabling LLMs to develop complex reasoning strategies and significantly improving capability-per-compute efficiency.

“Authored a 2025 year-in-review article analyzing paradigm changes in LLMs, including the emergence of Reinforcement Learning from Verifiable Rewards (RLVR).”

Reinforcement Learning has evolved from early psychology and mathematical foundations through the Deep RL revolution of the 2010s to modern applications like RLHF and GRPO in LLMs, with future potential still largely untapped.

“Discussed the future of reinforcement learning and stated that RL is terrible but everything before it was much worse; used FP16 precision format for nanochat.”

Leading AI researchers including Ilya Sutskever and Andrej Karpathy are revising AGI timelines and expressing skepticism about transformer scaling limits and LLM agent capabilities, raising doubts about current AI business model sustainability.

“Previously featured on a podcast discussing AI development and the future of AI systems.”

LLM Council is an open-source local web application that aggregates multiple LLMs into a council structure where models review each other's responses and a chairman model produces a final synthesized answer.

“Creator of the LLM Council project, a local web app that aggregates multiple LLMs for comparative analysis”

Richard Sutton's 'Bitter Lesson' argues that general search and compute methods outperform domain expertise, but engineers remain critical for properly framing search problems—as exemplified by tinygrad's approach to hardware optimization.

“Created micrograd which inspired George Hotz to build tinygrad, and appeared on Dwarkesh Patel's podcast calling the 2020s the 'decade of agents'.”

AGI timelines are realistically 5-10 years away but require significant work in integration, physical world sensors, safety, and research beyond current LLM capabilities, despite recent progress.

“Discussed AGI timelines, predicting a 'decade of agents' with 5-10X more pessimistic timelines than typical SF AI predictions, and wrote thoughts on BERT diffusion.”

Nanochat enables training GPT-2 capability models in 3 hours for ~$72 on a single 8XH100 GPU node, compared to $43,000 in 2019, with a complete open-source stack covering tokenization, pretraining, fine-tuning, and inference.

“Introduced nanochat, an open-source experimental harness for training LLMs that covers tokenization, pretraining, finetuning, evaluation, inference, and chat UI.”

Vibe coding—using AI as a high-level programming language—works initially but degrades as project context grows, requiring lightweight context management tools like Shallot for sustainable scaling.

“Originally coined the term 'vibe coding' in a viral tweet describing giving in to the vibes and forgetting the code even exists.”

Kairos, an AI-powered reading companion, enables fundamentally new ways to read by integrating AI directly into the reading experience, allowing readers to actively probe and discuss texts rather than passively consume them.

“Provided inspiration through an X post that led to the development of Kairos AI reading tool.”

Silicon Valley startups and major AI labs are racing to build reinforcement learning environments for training AI agents, with Anthropic reportedly planning to spend over $1 billion on the technology.

“Backer of Prime Intellect startup focused on RL environments”

Context engineering—properly balancing intent, environment, and conversation data—is the critical differentiator between mediocre and exceptional AI code review agents.

“Commented that good context engineering is what sets AI apps apart in response to discussions about context engineering for AI.”