World Models in 2026: Genie 3 at 24 fps, Cosmos 3 at 64B, V-JEPA 2 Planning in 16 Seconds, and an $8.2B Exit to AMD
What a world model is in 2026 and what actually shipped: Genie 3 at 720p and 20 to 24 fps for a few minutes behind a $249.99 subscription, World Labs' Marble and Atlas and the $8.2 billion all-stock sale to AMD, NVIDIA Cosmos 3 at 64B, 16B and 4B under OpenMDW 1.1, V-JEPA 2 planning a robot action in 16 seconds against Cosmos's four minutes, LeCun's AMI Labs raising $1.03 billion at $3.5 billion, the Physics-IQ ceiling of 29.5 out of 100, what inference costs, and a V-JEPA 2 embedding script I ran on a laptop CPU.
14 minOct 5, 2026Model Collapse in 2026: The Loop Nobody Runs, a 40% Synthetic Phi-4, a 49.9% AI Web, and 300T Human Tokens Left
What the model-collapse research tested versus the headline: Shumailov's Nature loop deletes real data every generation and takes OPT-125m from 20 to 28 perplexity, while accumulating data bounds the error; my 100-generation Gaussian experiment shows median variance falling from 0.98 to 0.36 under replacement and holding at 0.98 under accumulation; how Phi-4 (40% synthetic), Llama 3, DeepSeek-R1 (800k samples), Qwen3, Kimi K2 and Nemotron-CC-v2 (38%) actually use it; Epoch's 300T-token stock, Graphite's 49.9% AI web, Cloudflare's 38,065:1 crawl ratio and the $1.5B Bartz settlement.
15 minOct 5, 2026MLOps in 2026: Evals Are the New Unit Tests, Judges Agree With Experts 66% of the Time, and OpenAI's Evals API Shuts Down on 30 November
The 2026 state of MLOps, checked against primary sources: OpenAI is closing its Evals API (read-only 31 October, off 30 November) and winding down fine-tuning while buying Promptfoo; Langfuse went to ClickHouse and W&B to CoreWeave; GDPval's model grader agrees with experts 66% of the time against 71% for humans; ARC-AGI-3 went from 0.51% to 99.9% in five months; Terminal-Bench retires a task when every frontier model solves it 5 of 5 times; MLflow 3.16, Kubeflow 26.03, Feast 0.66, lakeFS under BSL, and a pytest gate on a Wilson lower bound that I actually ran.
14 minOct 5, 2026Agent Memory in 2026: Vendors Claim 92.5 on LoCoMo, a Plain Filesystem Scored 74, and Every Memory Product Lost to BM25
Agent memory systems in 2026, kept honest: Mem0's 66.9 on LoCoMo in its own paper versus 92.5 on its managed platform a year later, Zep's 75.1 rebuttal and the three bugs behind Mem0's 66.0 for Zep, Letta's 74.0 with only a filesystem, MemoryAgentBench putting every commercial system under 45 on retrieval against BM25 at 60.5, Hindsight's 89.6, first-party memory from ChatGPT, Claude and Gemini with dates, both labs' compaction APIs, what re-reading 115k tokens costs at cache-read prices, a Postgres and pgvector layer with supersession, and why memory poisoning is still unfixed.
13 minSep 9, 2026On-Device Inference in 2026: 160 Tokens a Second on a Laptop, a 2-Bit Model in Every iPhone, and the Bandwidth Arithmetic That Decides What Stays Local
What actually runs on the device now: Apple's 3B model trained at 2 bits and the WWDC26 framework that lets any model plug in, Gemini Nano inside AICore, Windows ML and the quiet replacement of Phi Silica, the NPU TOPS numbers and why they are the wrong metric, the small-model field from Gemma 4 E2B to Qwen3.5-2B to LFM2.5 with licences, measured decode speeds from Google's own tables and an independent phone benchmark, the runtimes (llama.cpp with multi-token prediction, LiteRT-LM, ExecuTorch 1.0, MLX), the memory-bandwidth arithmetic that predicts all of it, and the model-extraction attacks nobody has fixed. With commands to run one today.
14 minSep 9, 2026AI for Science in 2026: One Phase III Trial, Two Operational Weather Models, and Erdős Proofs Nobody Wants to Read
What has actually been validated outside the lab that made the claim: rentosertib's Phase III start, ECMWF and NOAA running AI forecasts operationally, Evo 2's viable phage genomes, the Erdős-problem scoreboard and its rediscovery problem, the GNoME and MatterGen materials that turned out not to be new, and the licence and pip-install status of every open model an engineer can run this week.
16 minSep 8, 2026Open Weights in 2026: Chinese Labs Ship the Frontier, the Licenses Grew Clauses, and Brussels Started Enforcing
DeepSeek V4, Kimi K3, GLM-5.3, Qwen3.8 and Xiaomi's MiMo are the open frontier now; Meta went closed and came back; the licences quietly acquired revenue thresholds; the EU AI Office got its enforcement powers on 2 August. What I found when I read every licence file, with the Hugging Face API calls that check them.
16 minSep 8, 2026Robot Foundation Models in 2026: 500,000 Hours of Human Video, 50-Step Action Chunks, and a Winning Score of 0.26
What vision-language-action models actually do, from π0's flow-matching action head to Gemini Robotics 2 and GR00T N1.7; who shipped humanoids (Unitree 5,500, AgiBot 15,000) and who didn't (Tesla, zero); the teleoperation behind the demos; and the benchmark numbers that say a home robot is not close. With a LeRobot dataset I loaded this morning.
14 minSep 8, 2026Agent Observability in 2026: The Spans Are Standard, the Standard Isn't Stable, and Nobody Reads the Logs
OpenTelemetry's GenAI conventions moved to their own repo in June and still carry the Development badge; every vendor from Datadog to Grafana ingests them anyway. I instrumented an agent loop with the official Python library, show the exact spans it emits, and go through what the 2026 eval papers say actually catches failures: pass^k, log inspection, and a step budget.
13 minSep 7, 2026LLM Inference in 2026: Prefill Is a Different Machine Now, and the KV Cache Has a Disk Tier
vLLM 0.28, SGLang on 400,000 GPUs, NVIDIA Dynamo 1.0 and llm-d in the CNCF all converged on the same architecture: split prefill from decode, route on cached prefixes, page the KV cache to CPU and SSD, and run MoE experts wide across a rack. What changed this year, with the papers, the flags, and the prices.
15 minSep 6, 2026Neuromorphic Computing in 2026: Real at Microwatts, Not for LLMs
Intel archived Lava and went quiet. BrainChip shipped production silicon. Sandia runs a billion neurons. An infrastructure engineer's honest read of where spiking hardware actually wins, with benchmark numbers and a runnable SNN.
12 min