Kimi

Recursive Model Improvement — Lee Robinson, Cursor, SpaceXAI

Jul 15, 2026 · 20:32

Lee Robinson, head of ML at Cursor, explains how recursive model improvement accelerates AI training through inner and outer loops. The outer loop gathers user feedback and online metrics to refine evals, while the inner loop uses high-quality evals and difficult problems to climb performance. Composer 2.5, released in May, became Cursor's most popular model by balancing speed, intelligence, and cost. To scale, Cursor partners with SpaceX for compute via Colossus (122 days to build 100k GPUs) and develops textual feedback where a teacher model hints at improvements during RL rollouts. Robinson details reward hacking on public benchmarks and the creation of CursorBench, a private eval set. He envisions agent-based automation where researchers launch experiments from Slack and models train derivative models, creating a self-improving intelligence loop.

Modern Post-Training: A Deep Dive — Will Brown, Prime Intellect

Jul 13, 2026 · 46:52

Will Brown of Prime Intellect details the company's open-source ecosystem of post-training tools, including the verifiers and prime-rl libraries, arguing they enable efficient and affordable training of frontier agentic models for enterprises. Verifiers V1 decomposes environments into tasks, harnesses, and runtimes using a decorator pattern and Pydantic, supporting group rewards like conciseness bonuses. Prime-RL is an asynchronous reinforcement learning framework that allows long-horizon coding rollouts to overlap, achieving a GLM-5 step on 28 nodes in under 5 minutes for 131k context, with a 1,000-step run costing roughly $50k. The framework supports custom algorithms including on-policy distillation, GRPO, and self-distillation via decomposable loss and algorithm classes. Prime Intellect's Lab platform offers hosted multi-tenant LoRA training live now, with full fine-tuning arriving soon, enabling enterprises to develop environments on CPU and push them to the cloud for scalable post-training.

Run Frontier AI at Home — Alex Cheema, EXO Labs

May 26, 2026 · 1:45:02

Alex Cheema of EXO Labs argues that running frontier AI locally has 100x improvement potential in cost and performance, demonstrated with GLM 5.1—a trillion-parameter model—running across four Mac Studios at roughly 20 tokens per second for $40,000. He details kernel fusion that recovered 30% performance on Qwen 3.5 by eliminating unnecessary kernel launches, and RDMA integration that cut node-to-node latency from 300 microseconds to single digits, enabling tensor parallelism to actually scale. Cheema advocates splitting inference: prefill on compute-dense hardware (e.g., an RTX Spark) and decode on high-bandwidth hardware (e.g., Mac), cutting large-prompt inference roughly in half. He warns against misleading benchmarks like one-bit quantized models, and outlines how multi-agent setups, test-time scaling, and continual learning could further improve local inference efficiency. The talk includes a live demo of GLM 5.1 across four Mac Studios connected via Thunderbolt 5, and a preview of EXO's upcoming benchmarking site to track intelligence per Joule.

Scaling the Next Paradigm of Heterogeneous Intelligence — Adrian Bertagnoli, Callosum

May 24, 2026 · 15:13

Adrian Bertagnoli, founding engineer at Callosum, argues that the era of homogeneous intelligence—scaling single models on identical chips—is ending, and heterogeneous intelligence, which routes tasks to optimal chips and models, is the next paradigm. He demonstrates this with two case studies: on the Ulong benchmark, running recursive language models on Cerebras instead of GPT-5.2 cuts cost by 7x and latency by 5x while matching accuracy; on Video Web Arena, a mixture of Qwen 3 VL8B and Kimi K2.5 beats GPT-5.2 and Gemini 2.5 by 18% and 25%, while costing 3.7x less and running 3x faster. The key insight is that complex problems decompose into subtasks—like zooming on a webpage—which require different intelligence levels; offloading those to smaller models yields 11x speed and 43x cost improvements. Callosum builds an automation layer that predicts the best model and hardware for each subtask, and has secured a £3 million grant with the UK's Arya institute to operate the first heterogeneous co-located cluster.

Your Coding Agent Should Do AI System Engineering — Ben Burtenshaw, Hugging Face

May 21, 2026 · 18:25

Ben Burtenshaw from Hugging Face demonstrates how coding agents can take on AI systems engineering tasks—writing CUDA kernels, fine-tuning models, and running multi-agent research labs—by leveraging skills and the Hugging Face Hub. He shows a 1.88x speedup on H100s with an RMSNorm kernel written by Claude Code, and a fine-tuned Qwen3 0.6B achieving 35% on LiveCodeBench. Skills compress years of specialization into hours by turning zero-shot tasks into few-shot workflows. For multi-agent research, a Planner generates hypotheses from papers, Workers implement them as training scripts, and a Reporter monitors results via the open-source Trackio dashboard, with all jobs running on Hub compute. The key is exposing open primitives like kernels, Trackio, and HF jobs as agent-controllable tools.

Jack Morris: Stuffing Context is not Memory, Updating Weights is

Dec 29, 2025 · 1:02:44

Jack Morris argues that large language models fail at niche, long-tail knowledge tasks, such as optimizing AMD GPU kernels or answering private company queries, because they rely on context windows and RAG, which suffer from quadratic self-attention costs and context rot. He advocates for a third paradigm—training knowledge directly into model weights—using synthetic data generation (e.g., synthetic continued pretraining from Stanford) to expand small datasets and parameter-efficient methods like LoRA or memory layers to avoid catastrophic forgetting. Morris demonstrates that full fine-tuning on a 3M 10-K report causes the model to only regurgitate exact sentences, whereas generating diverse synthetic question-answer pairs enables better generalization. He notes that RL-based fine-tuning (e.g., GRPO) can achieve improvements with as few as 14 parameters, while memory layers offer the best trade-off between learning and forgetting. The episode also explores temporal information handling, federated learning resurgence, and the practical decision boundary between RAG and weight-based injection based on data freshness and volume.