[(Kevin Buzzard says) This is a guest post by Mitchell Taylor, a mathematician who gave a talk at Imperial’s PDE seminar a few weeks ago on the topic below.
Latest in Machine Learning
author · site-
Lean for the working mathematician
-
The Cognitive Gym
Why the Best Way to Learn With AI Is to Make It Harder
-
The US-China Open-Weight Feedback Loop
How US and Chinese open-weight LLMs build on each other, illustrated with a timeline of architecture ideas from 2021 to 2026.
-
Sidework plugin v0.1: Monitor open source projects
When I get back to an open source project, I want to know which pull request needs my attention and why.
-
Safety stock when the supplier is late, worked with numbers
Safety stock = z·√(L·σd² + d²·σL²), worked for one product: why lead-time spread dominates, why the two buffers do not add, and fill rate vs service level.
-
Introduction to Bayesian Optimization
An introduction to Bayesian optimization for expensive black-box functions.
-
MiniMax H3 Image Edit Recipes: I Tested Whether the JSON Prompt Format Matters
Three MiniMax H3 Image edits that people share as ComfyUI node graphs — a character placed in a scene from two references, clothes extraction, a three-view character sheet — rebuilt as one YAML file on a single RTX 4090.
-
Better Chatters Than Writers
If I’m in a hurry on a story, I might ask you to email me a quote.
-
Neural Network Regression with Arctan Hidden Node Activation Using C#
Neural network regressors have two activation functions.
-
New video for Edge SpAIce project
Check out this fantastic video summarising the Edge SpAIce project!
-
Psychological Safety Is the Competitive Advantage
Psychological safety is not a soft corporate perk; it is the biological prerequisite for high performance.
-
“The AI Trust Problem: Risks and Business Opportunities” on the Pure AI Web Site
I contributed some technical content and opinions to an article titled “The AI Trust Problem: Risks and Business Opportunities” in the November 2026 edition of the Pure AI web site.
-
Training LoRAs for MOSS-Image and Qwen-Image 2.1 on One RTX 4090: 24 Photos, Under an Hour, ~12 GB of VRAM
I trained the same 24-photo subject LoRA on MOSS-Image Base and Qwen-Image 2.1 on a single 24 GB RTX 4090, sampling every 250 steps with fixed seeds.
-
Who changed the AI agent’s rules?
Image: illustration made with Claude for this post We spend a lot of time measuring what AI coding agents produce.
-
MiniMax H3 Image on One RTX 4090: 4 MP in 15 Seconds, Perfect Typography, and an Edit Model That Doesn't Chain
MiniMax's H3 Image is a 20B, 6-step, natively 4-megapixel image model with a 32B text encoder.
-
OmniLLL Part 1: Discussion #86 Post-Mortem & The Zero-Python Pure-Rust Awakening
What Siya68 and DavidAU's Discussion #86 actually proved on 16 GB RX 6900 XT, why OmniLLL eliminates Python for #![no_std] p/q Rust + Metal/CUDA/C64, and how M1 Max 64GB fits 27B multimodal training with 26-33 GB headroom.
-
Pi Durable ♥ Actors: One Agent, One Owner, Any Worker
I have always thought of agents as actors with, well, agency.
-
MAPE picked the forecast that RMSE called worse
MAPE formula worked on five days: 8.65% vs 7.14%, the reverse of RMSE.
-
Attention in the y = abx Toy Model
Learning dynamics of a scalar linear attention layer on in-context regression
-
From a 27-Hour HNSW Index to 3 Minutes: Upgrading pg_bio with Tiger Data’s pg_vectorsearch (PRISM), RaBitQ 1-Bit Error Bounds, and 1.2ms Dark Proteome Homology
How combining Tiger Data's brand-new open-source pg_vectorsearch extension (PRISM index access method with 1-bit RaBitQ error bounds and SOAR routing) with pgvector halfvec(1280) STORAGE PLAIN and pg_bio's Rust Smith-Waterman operator shrank our 2.34-million-protein index by…
-
Crossing 67.5 Million Dense Edges in PostgreSQL 18: Pass-3 Ring Checkpointing, a 128x Faster Global Fold Gate, and Solving DUF2071, DUF1156 & DUF7382
How idempotent Pass-3 HNSW ring checkpointing (k = 25..36) propelled our PostgreSQL 18 discovery swarm from 20.31M to 67.50 million dense edges (+47.19M new edges), how a biophysical Global Fold & Length-Consistency Gate stopped 56-residue single-helix fragments from hijacking…
-
Design workflows for the agentic age
Photo by Kristina Alexanderson (internetstiftelsen), CC-BY-SA I've been having some really interesting discussions, of late, with colleagues (you know who you are) about agentic workflows.
-
NFL 2026 Week 11 Predictions – Zoltar Likes the Broncos to Destroy the Raiders
Zoltar is my NFL football prediction computer program.
-
Verifiable Encryption: Escrowing a Crypto Wallet Key Without Revealing It
Alice keeps her savings in a crypto wallet, and wants a loan from her bank using them as collateral.
-
Introduction to Gaussian Process Regression
Derives Gaussian process regression from joint Gaussian distributions and conditioning.
-
Introduction to TrueSkill: Learning from Game Results
Builds the classic TrueSkill model from Gaussian skill beliefs and noisy match performance.
-
Pre-Training Without Learning Rate Decay
A quick comparison of LLM pre-training recipes using cosine decay, warmup-stable-decay, or constant learning rates with checkpoint averaging.
-
How We Built the AI Factory: Workspaces, Runners and Reports
How the AI Factory is designed on top of Multica: one workspace per team, agents whose access does the locking, and a runner per workspace inside a box we control.
-
I Use AI to Improve My Random Forest Regression System Using JavaScript
A simple decision tree for regression will almost always overfit the training data where the model accuracy on the training data is near 100% but accuracy on new, previously unseen data is very poor.
-
AI vs Programmers
AI vs Programmers This post is about my use of local AI systems for coding tasks.
RSS for this view · OPML of the feeds with posts in this view (184 feeds)